| Internet-Draft | AGTP | May 2026 |
| Hood | Expires 13 November 2026 | [Page] |
- Workgroup:
- Independent Submission
- Internet-Draft:
- draft-hood-independent-agtp-07
- Published:
- Intended Status:
- Informational
- Expires:
Agent Transfer Protocol (AGTP)
Abstract
AI agents and agentic systems generate a growing volume of intent-driven, unstructured, and undifferentiated traffic that flows through HTTP indistinguishably from human-initiated requests. HTTP lacks the semantic vocabulary, observability primitives, and identity mechanisms required by agent systems operating at scale. Existing protocols described as Agent Group Messaging Protocols (AGMP), including MCP, ACP, A2A, and ANP, are messaging-layer constructs that presuppose HTTP as their transport. They do not address the underlying transport problem.¶
This document defines the Agent Transfer Protocol (AGTP): a dedicated
application-layer protocol for AI agent traffic. AGTP is a runtime
contract negotiation substrate (RCNS): a transport that fixes only a
twelve-method protocol floor and negotiates any additional method
surface at runtime between agent and server in a single round-trip,
governed by the AGTP-API companion specification [AGTP-API], which
defines the curated method catalog, path grammar, endpoint primitive, and
synthesis semantics. Version 07 confirms the
IANA-registered agtp:// URI scheme and IANA-assigned port 4480 for
TCP/TLS and QUIC, formalizes Form 1a URI grammar
(agtp://{agent-id}@{host}) for direct addressing, renames the Agent
Manifest Document to the Agent Identity Document with an enumerated
schema, redesigns the protocol-defined method floor to a 12-method set
organized as six cognitive verbs (QUERY, DISCOVER, DESCRIBE, SUMMARIZE,
PLAN, PROPOSE) and six mechanics verbs (EXECUTE, DELEGATE, ESCALATE,
CONFIRM, SUSPEND, NOTIFY), establishes AGTP as a substrate for
higher-level agent frameworks (MCP, A2A, ACP) carried as content types
inside AGTP method invocations, renumbers AGTP-specific status codes out
of HTTP-assigned space to avoid semantic collision, mandates explicit
Content-Length framing with a prohibition on TLS socket-level half-close,
adds a .well-known/agtp bootstrap convention per RFC 8615, deprecates
the AGIS reference and the proposed AGTP-Methods specification by
folding both into the unified AGTP-API contract layer, adds status
codes 405 (Method Not Allowed), 459 (Method Violation), and 460
(Endpoint Violation) per the AGTP-API contract model, and adopts
"Agent Genesis" as the canonical term for the permanent signed
origin document. Version 06 prepared
the IANA Service Name and Port Number application and consolidated the
URI scheme registration. Version 05 restored the canonical Agent-ID as
the primary identity primitive and decoupled Trust Tier 1 verification
from DNS as a sole requirement. A canonical Agent-ID is derived from the
agent's Agent Genesis hash and is authoritative in every AGTP
protocol operation. Three equivalent verification paths are recognized
for Trust Tier 1: DNS-anchored verification via RFC 8555 ACME challenge,
log-anchored verification via Agent Genesis inclusion in an
append-only transparency log aligned with RFC 9162 and RFC 9943 (SCITT),
and hybrid verification combining DNS control with blockchain address
ownership. Version 04 introduced
normative integration hooks for the AGTP Merchant Identity and Agentic
Commerce Binding specification [AGTP-MERCHANT], which defines the
merchant-side identity model that complements AGTP's agent-side
identity model. AGTP SHOULD prefer QUIC for new implementations and
MUST support TCP/TLS for compatibility and fallback. It is designed to
be composable with existing agent frameworks, not to replace them.¶
Status of This Memo
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 13 November 2026.¶
Copyright Notice
Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document.¶
1. Introduction
Note Regarding Intellectual Property: Implementers should be
aware that extensions and certain mechanisms referenced in this
document -- including the Agent Certificate extension (Section 7.2),
the ACTIVATE method, the Agent Genesis mechanism
(Section 5.7), and the .agent and .nomo file format
specifications (Section 2) -- may be subject to pending patent
applications by the author. The core AGTP specification is intended
for open implementation without royalty obligation. The licensor is
prepared to grant a royalty-free license to implementers consistent
with [RFC8179]. IPR disclosures:
https://datatracker.ietf.org/ipr/ -- see also Section 7.7.¶
1.1. Background
The deployment of AI agents and multi-agent systems is accelerating across enterprise, research, and consumer contexts. These systems execute complex, multi-step workflows, querying data sources, booking resources, delegating subtasks to peer agents, and escalating decisions to human principals, with minimal or no human supervision per transaction.¶
Unlike human-initiated web traffic, agent-generated traffic is dynamic, high-frequency, intent-driven, and often stateful across sequences of related requests. The infrastructure carrying this traffic was not designed with these properties in mind.¶
1.2. Limitations of HTTP for Agent Traffic
HTTP has served as the internet's primary application-layer transport for over three decades. Its evolution through HTTP/2 [RFC7540] and HTTP/3 [RFC9114] has improved performance, multiplexing, and latency. However, the fundamental model of HTTP being stateless, resource-oriented, human-initiated request/response, creates specific failures when applied to agentic systems at scale:¶
-
Traffic indistinguishability: Agent-generated requests are structurally identical to human-initiated requests at the transport layer. Operators cannot identify, route, or govern agent traffic without application-layer instrumentation.¶
-
Method vocabulary mismatch: HTTP's method set (GET, POST, PUT, DELETE, PATCH) describes resource operations. Agent traffic expresses purposeful intent, summarize, book, delegate, escalate. The mismatch forces intent into request bodies, invisible to protocol-level handlers.¶
-
Identity and attribution absence: HTTP carries no native mechanism for asserting agent identity, declared authority scope, or the principal accountable for an agent's actions.¶
-
Session semantics mismatch: HTTP's stateless model is optimized for isolated request/response cycles. Agent workflows are inherently stateful sequences.¶
1.3. Why Not Evolve HTTP?
A natural question is whether these limitations could be addressed by extending HTTP rather than defining a new protocol. There are three specific reasons why HTTP extension is not the preferred path.¶
First, the HTTP method registry is effectively frozen for new semantics. [RFC9110] defines the HTTP method registry with IETF Review as the registration procedure, meaning new methods require a full IETF consensus process and must be backward-compatible with existing HTTP implementations. Adding intent-based verbs (SUMMARIZE, DELEGATE, ESCALATE) to HTTP would require every HTTP client, server, proxy, and middleware component to ignore or handle unknown methods gracefully, a compatibility constraint that limits how agent-specific semantics can be expressed at the protocol level.¶
Second, HTTP carries decades of backward-compatibility constraints. Features such as persistent agent identity headers, authority scope declarations, and session-level governance semantics would require HTTP extensions that interact unpredictably with existing caching, proxy, and CDN behavior designed for human-generated traffic patterns.¶
Third, the observability goal making agent traffic distinguishable from human traffic at the infrastructure layer cannot be achieved by adding fields to HTTP. Infrastructure components route and filter HTTP traffic based on methods and headers that are identical across agent and human requests. A protocol-level separation is necessary to give infrastructure the signal it needs.¶
AGTP is therefore designed as a dedicated protocol rather than an HTTP extension. HTTP and AGTP coexist: human traffic continues to flow over HTTP; agent traffic flows over AGTP. The two protocols serve different classes of network participant.¶
Note: The abbreviation AGTP is used in this document to distinguish the Agent Transfer Protocol from the Authenticated Transfer Protocol (ATP) working group currently chartered within the IETF. The URI agtp:// is proposed for IANA registration as a new and distinct scheme.¶
1.4. Motivation for a Dedicated Protocol
These limitations are architectural, not implementational. They cannot be resolved by better middleware or application code layered on HTTP. They require a protocol designed from first principles for AI agent systems.¶
AGTP is that protocol. It provides a dedicated transport environment for agent traffic with: native intent-based methods, mandatory agent identity headers, protocol-level authority scope declaration, and a status code vocabulary for the conditions AI systems encounter.¶
1.5. Scope and Target Audience
This document covers AGTP architecture, design principles, stack position, request and response header format, agent-native method definitions and semantics, status code vocabulary, security considerations, and IANA considerations.¶
The Agent Certificate extension for cryptographic binding of agent identity to AGTP header fields is described at a high level in Section 7.2. Full specification is provided in a separate companion document: [AGTP-CERT]. That extension may be subject to pending intellectual property claims; see Section 7.7 and the IPR Notice preceding the Abstract.¶
Merchant-side identity verification for PURCHASE counterparties is
described at a high level in Section 8 of this document and specified
in full in a separate companion: [AGTP-MERCHANT]. This document
registers the merchant-related request headers, the 458 Counterparty
Unverified status code, and the merchant and intent Authority-Scope
domains; the Merchant Manifest Document, Merchant Agent Genesis,
counterparty verification procedure, and Intent Assertion JWT format
are specified in the companion.¶
Target audience: AI agent developers, protocol designers, cloud and network infrastructure providers, enterprise security and compliance architects, and standards community participants.¶
1.6. AGTP as the Transport Foundation for Agent Group Messaging Protocols
AGTP is the purpose-built transport and governance layer for Agent Group Messaging Protocols (AGMPs): the category of higher-layer AI agent messaging standards that includes the Model Context Protocol (MCP) [MCP], the Agent-to-Agent Protocol (A2A) [A2A], the Agent Communication Protocol (ACP) [ACP], and emerging others.¶
AGMPs define what agents say. AGTP defines how those messages move, who sent them, and under what authority. AGTP provides the narrow-waist foundation that AGMPs inherit without modification: intent-native methods, mandatory agent identity and scoping, resource budget enforcement, observability hooks, and normative composition profiles. A deployment running any AGMP over AGTP gains transport-level governance without changes to the messaging layer.¶
The AGMP category term is introduced in this document to provide a stable collective reference for the class of protocols that AGTP serves as substrate. It is not a formal IETF term of art; it is a descriptive classification. Individual AGMP specifications retain their own names and development paths. AGTP does not govern, modify, or supersede any AGMP.¶
+-----------------------------------------------------+ | Agent Application Logic | +-----------------------------------------------------+ | AGMP Layer: MCP / A2A / ACP / ANP [optional] | +-----------------------------------------------------+ | AGTP - Agent Transfer Protocol [this spec] | +-----------------------------------------------------+ | TLS 1.3+ [mandatory] | +-----------------------------------------------------+ | TCP / QUIC / UDP | +-----------------------------------------------------+
2. Terminology
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals.¶
- Agent:
-
An AI software system that executes tasks, makes decisions, and takes actions without continuous human supervision per transaction.¶
- Principal:
-
The human, organization, or system that authorized an agent to act and is accountable for its actions.¶
- Agent-ID:
-
A unique identifier for a specific agent instance. Carried in the
Agent-IDrequest header on non-anonymous AGTP requests, and in theagent_idfield of the Agent Identity Document.¶ - Principal-ID:
-
The identifier of the principal on whose behalf an agent operates. Carried in the agent identity document referenced by
Agent-ID; not transmitted as a separate request header.¶ - Authority-Scope:
-
A declared set of permissions defining what actions an agent is authorized to take, in the format
domain:actionordomain:*. Declared in the agent's identity document. MAY be carried on individual requests as a claimed-scopes header narrowing the agent's full authorized set to those needed for the request; claimed scopes MUST be a subset of the document's declared set.¶ - Intent Method:
-
An AGTP method name expressing the agent's purpose, as distinguished from HTTP resource-operation verbs.¶
- Delegation Chain:
-
An ordered record of Agent-IDs representing the sequence of delegations that produced the current request.¶
- Escalation:
-
An agent's intentional deferral of a decision or action to a human principal or higher-authority agent.¶
- Attribution Record:
-
A logged record of an agent action sufficient for audit and compliance purposes.¶
- Session:
-
An AGTP persistent connection context shared across multiple method invocations within a single agent workflow.¶
- SEP (Scope-Enforcement Point):
-
An AGTP-aware infrastructure component, load balancer, gateway, and proxy, that enforces Authority-Scope compliance without application-layer access. Requires the Agent Certificate extension ([AGTP-CERT]).¶
- Agent Package (.agent):
-
A portable, open deployment artifact for an AI agent. An
.agentfile contains an embedded Agent Manifest, an integrity hash covering all package contents, and a behavioral trust score computed at packaging time. The.agentformat is an open specification. It is analogous to a container image: a self-describing, portable unit of deployment. The.agentsuffix is a file format designator and MUST NOT appear as a hostname component or top-level label inagtp://URIs. Note: the.agentfile format specification may be subject to pending patent claims by the author; see Section 7.7.¶ - Governed Agent Package (.nomo):
-
A deployment artifact in the
.nomoformat, which extends the.agentformat with a CA-signed certificate chain binding the package to a verified governance zone and issuing principal. The.nomoformat is to.agentas HTTPS is to HTTP: the same structural foundation with an added layer of cryptographic trust. A.nomopackage is required for agents operating at Trust Tier 1 (see Section 5.2). The.nomosuffix is a file format designator and MUST NOT appear as a hostname component inagtp://URIs.¶The name derives from the Greek nomos (νόμος), meaning law, rule, or governance, the same root that underlies autonomy (self-law), nomocracy (rule of law), and onomastics. A
.nomopackage is literally an agent operating under law: its behavior is bounded by a cryptographically enforced governance context at the packaging layer. Note: the.nomofile format specification may be subject to pending patent claims by the author; see Section 7.7.¶ - Agent Transfer Document (.agtp):
-
The wire-level manifest document format defined by this specification. An
.agtpdocument is a signed JSON structure containing the fields defined in Section 5.5 (Agent Identity Document). It is the output format returned by all AGTP URI resolution requests. Both.agentand.nomopackages produce.agtpdocuments when queried; the.agtpformat is the protocol's canonical representation of agent identity and is independent of the underlying packaging format. The.agtpsuffix MAY appear in filenames for stored manifest documents but MUST NOT appear inagtp://URIs. The Content-Type for.agtpdocuments isapplication/vnd.agtp+json.¶ - URI (AGTP):
-
An
agtp://scheme URI that identifies an agent or agent namespace. AGTP URIs are addresses, not filenames. File extensions (.agent,.nomo,.agtp) MUST NOT appear in canonical AGTP URIs. See Section 5.1 for the canonical URI forms and resolution semantics.¶ - Agent Namespace Document:
-
A cryptographically signed
application/vnd.agtp+jsondocument returned in response to a request targeting an organization's agent registry root (e.g.,agtp://acme.tld/agents). Lists all Active agents registered under the organization's governance zone. The document is generated and re-signed by the governance platform on any registry change. It is not a manually editable file. See Section 5.4.¶ - Agent Identity Document:
-
A cryptographically signed
application/vnd.agtp+jsondocument returned in response to a request targeting a specific agent (e.g.,agtp://acme.tld/agents/customer-service). Contains the agent's Agent Genesis fields, lifecycle state, behavioral trust score, authority scope categories, supported methods, and governance zone. Derived directly from the agent's.agentor.nomopackage; the package integrity hash is verified before the manifest is served. See Section 5.5.¶ - Agent Genesis:
-
A cryptographically signed origin document issued to an agent at registration time by a governance platform. The Agent Genesis is the genesis record of an agent's existence: it establishes the agent's identity, ownership, authorized scope, behavioral archetype, and governance zone before the agent takes any action. Authority is issued through the Agent Genesis; it is never self-assumed.¶
The Agent Genesis is the source document from which the Agent Identity Document (Section 6.4) is derived when an AGTP URI is resolved. The
certificate_hashfield of the Agent Genesis is the basis for the agent's canonical Agent-ID. In this sense the Agent Genesis functions as the agent's permanent origin record: issued once at creation, permanently bound to the agent, and the authoritative identity record from which all other identity representations derive.¶Agent Genesis fields map to AGTP protocol elements:
agent_idmaps to theAgent-IDheader on every request;owneris the principal identifier recorded in the agent identity document referenced byAgent-ID(not a separate header);scopeis the Authority-Scope set declared in the agent identity document, optionally narrowed per-request via theAuthority-Scopeheader. See Section 5.7.¶Anonymous agents are ungovernable. Without an Agent Genesis, there is no mechanism to trace decisions to a responsible principal, enforce scope boundaries, or maintain a meaningful audit trail.¶
The taxonomy is: Agent Genesis (the permanent signed governance- layer origin document) → canonical Agent-ID (the 256-bit hash of the Agent Genesis, used in all AGTP protocol operations) → Agent Certificate (an optional X.509 v3 credential for TLS mutual authentication; specified in [AGTP-CERT]). Note: the Agent Genesis mechanism may be subject to pending patent claims by the author; see Section 7.7.¶
- Governance Token:
-
A signed, time-limited JWT artifact issued by a governance runtime that encodes a specific governance decision for a specific action. Governance tokens are the runtime companion to the static Agent Genesis: where the Agent Genesis establishes persistent identity, the Governance Token carries a bounded authorization for a single action or session. Tokens carry the governance verdict (ALLOW, DENY), the agent ID, action details, trust score dimensions, issuer identity, and expiry. Default TTL: 30 seconds. Tokens MUST NOT be reused across actions; each action requires a fresh evaluation and a fresh token.¶
- Trust Tier:
-
A classification assigned to an agent based on the strength of identity verification backing its registration. Tier 1 (Verified): org anchor is a real DNS domain with confirmed ownership and a
.nomogoverned package. Tier 2 (Org-Asserted): org label is present but DNS ownership is unverified;.agentpackage acceptable. Tier 3 (Experimental): X- prefix required; not discoverable through the public AGTP registry. See Section 5.2.¶ - AGMP (Agent Group Messaging Protocol):
-
The collective term for higher-layer AI agent messaging standards that operate over AGTP as their transport substrate, including MCP [MCP], A2A [A2A], ACP [ACP], and ANP [ANP]. AGMPs define what agents say to each other. AGTP defines how those messages move. The term is introduced in this document as a descriptive classification; it is not a formal IETF term of art.¶
- DESCRIBE:
-
An AGTP cognitive floor method that returns the declared capabilities, supported modalities, method vocabulary, and versioned feature set of a specific agent endpoint. Distinguished from URI resolution (which returns identity) by returning operational capability metadata suitable for pre-task negotiation. If the
capability_domainsparameter is omitted, the server SHOULD return all supported domains. Category: ACQUIRE.¶ - SUSPEND (method):
-
An AGTP mechanics floor method that places a specific active session workflow into a recoverable paused state, issuing a resumption nonce for re-entry. Distinguished from the lifecycle SUSPEND event (Section 6.7.6): method-level SUSPEND is session-scoped and does not affect the agent's registry lifecycle state or Agent Genesis validity. Category: ORCHESTRATE.¶
- Budget-Limit:
-
A request header declaring the maximum resource consumption the principal authorizes for a method invocation, expressed as comma-separated
unit=valuetokens drawn from the IANA AGTP Budget Unit Registry per [RFC9110] list-valued header conventions. Example:Budget-Limit: tokens=5000, compute-seconds=120, financial=10.00USD, ttl=3600. Exceeding the declared limit MUST cause the server to return 456 Budget Exceeded rather than continue execution. Note: ttl= is RECOMMENDED to bound budget lifetime. Reserved for v01+ per Section 5.5.¶ - AGTP-Zone-ID:
-
A request header declaring the network zone or organizational boundary within which a request must be processed. Scope-Enforcement Points (SEPs) MUST enforce zone boundaries and MUST return 457 Zone Violation if a DELEGATE request would route outside the declared zone.¶
3. Problem Statement
AGTP is motivated by three distinct, compounding failures in how current internet infrastructure handles AI agent traffic.¶
3.1. Problem 1: Undifferentiated Agent Traffic on HTTP
AI agents generate intent-driven, structured traffic that is functionally invisible to the infrastructure it traverses. This traffic flows through HTTP alongside human traffic with no protocol-level differentiation. Observability failure, routing inefficiency, and security blindness result, operators cannot determine what fraction of traffic is agent-generated without application-layer instrumentation that is expensive, inconsistent, and easy to circumvent.¶
AGTP response: a dedicated protocol environment for agent traffic. Infrastructure can distinguish, route, monitor, and govern agent traffic natively.¶
3.2. Problem 2: Semantic Mismatch Between Agent Intent and Available Methods
AI agents operate on intent. HTTP's method vocabulary was designed to describe operations on resources, not purposeful action. When an agent intends to SUMMARIZE a document, EXECUTE a reservation, and PLAN a sequence, all three arrive as POST requests. The server receives identical verbs with meaningfully different intent buried in request bodies, invisible to any protocol-level handler.¶
AGTP response: a vocabulary of agent-native methods that express intent at the protocol level.¶
3.4. Problem Summary
| # | Problem | Current Failure | AGTP Response |
|---|---|---|---|
| 1 | Undifferentiated traffic | HTTP cannot separate agent traffic | Dedicated protocol environment |
| 2 | Semantic mismatch | HTTP verbs obscure agent intent | Native intent-based method vocabulary |
| 3 | No protocol-level identity | Attribution is untraceable | Agent identity and scope in headers |
5. Protocol Overview
5.1. Stack Position
AGTP is an application-layer protocol. It operates above the transport layer (TCP, UDP, or QUIC) and is wrapped by TLS. It sits below any agent messaging protocol in deployments that use one.¶
-
SHOULD prefer QUIC [RFC9000] [RFC9001] for new deployments (lower latency, multiplexing without head-of-line blocking, 0-RTT connection establishment).¶
-
MUST support TCP/TLS as a fallback for compatibility with existing infrastructure.¶
-
MAY run over UDP where QUIC is not available, subject to implementor-defined reliability guarantees.¶
AGTP uses port 4480 (TCP and UDP), assigned by IANA under the
service names agtp (TCP/TLS) and agtp-quic (QUIC). The port
assignment is permanent and applies to all AGTP transports. Full IANA
registration metadata is documented in Section 11.1.¶
5.2. Design Principles
- Minimalist core:
-
The base spec defines only what is necessary for agent traffic differentiation, method semantics, and identity headers. Extensions belong in companion specifications.¶
- Extensible by design:
-
New methods are registered through an IANA-managed Method Registry. New header fields follow a defined extension convention. Additive changes do not require a version increment.¶
- Agent-native:
-
Every design decision assumes the initiating party is an AI system, not a human.¶
- Secure by default:
-
TLS 1.3 or higher is mandatory. Unencrypted AGTP connections MUST be rejected. Agent identity headers are present on every request.¶
- Observable by design:
-
Native metadata in every AGTP header provides the minimum information needed for routing, monitoring, and audit without application-layer instrumentation.¶
- Composable:
-
AGTP works alongside existing agent messaging protocols without requiring modification to those protocols.¶
- Runtime contract negotiation:
-
AGTP fixes the protocol surface at twelve methods. Beyond that floor, the surface is not predetermined. An agent that needs an endpoint the server does not advertise proposes the endpoint via PROPOSE with an AGTP-API endpoint definition; the server evaluates whether it can synthesize the endpoint from existing capabilities and either instantiates it as a session-scoped endpoint or refuses. The negotiation completes in a single round-trip. See Section 5.3.¶
5.3. AGTP as a Runtime Contract Negotiation Substrate
AGTP is a runtime contract negotiation substrate (RCNS). The substrate fixes a small floor of twelve protocol-level methods (the cognitive and mechanics verbs of Section 7.1); beyond that floor, the method and endpoint surface that any given AGTP server presents is not fixed in advance. It is negotiated at runtime between the agent and the server, in a single round-trip, governed by the AGTP-API companion specification [AGTP-API] which defines the curated method catalog, path grammar, endpoint primitive, semantic block, schemas, and synthesis semantics.¶
The negotiation loop is:¶
-
An agent encounters an AGTP server and inspects its endpoint surface via DESCRIBE or by retrieving the server manifest as defined in [AGTP-API].¶
-
If the endpoints the agent needs are not present, the agent submits a PROPOSE request carrying an AGTP-API endpoint specification: a verb (drawn from the AGTP-API approved verb list), a path (conforming to AGTP-API path grammar), a semantic block (intent, actor, outcome, capability classification, confidence guidance, impact tier, idempotency), input and output schemas, and declared error conditions.¶
-
The server evaluates the proposal against AGTP-API contract rules and against its own capability surface. The server determines whether the proposed endpoint can be synthesized from existing endpoints, what authority scope is required, and what governance constraints apply.¶
-
The server either instantiates the proposed endpoint as a session-scoped endpoint and returns 263 Proposal Approved with the AGTP-API endpoint definition, or refuses with 463 Proposal Rejected and a structured reason.¶
The full negotiation completes in a single PROPOSE → response round-trip. No registry update, no human-in-the-loop approval, no out-of-band coordination is required for an agent and a server to agree on a new endpoint, provided the endpoint is AGTP-API conformant and within the server's capability envelope.¶
This property distinguishes AGTP from protocols whose contract surface is fixed at design time. HTTP's method registry is effectively frozen (see Appendix D); proposing a new HTTP method requires IETF consensus over multi-year timelines. gRPC services expose a fixed service contract defined at compile time. Most agent frameworks built on HTTP inherit this constraint and work around it by overloading POST. AGTP treats the contract surface as dynamically negotiable, with AGTP-API as the contract layer that keeps negotiation safe.¶
The RCNS property is what makes composition with higher-level agent frameworks (Section 8) tractable. A server that wants to expose a framework-specific operation as a first-class AGTP endpoint does not need to wait for IANA registration; it negotiates the endpoint via PROPOSE at the moment an agent asks for it, governed by AGTP-API. The floor of twelve methods provides interoperability; AGTP-API provides contract safety; PROPOSE provides expressivity.¶
Implementations MAY choose not to participate in runtime negotiation. A server that supports only the twelve-method floor and returns 463 Proposal Rejected for every PROPOSE request is fully conformant with this specification. Runtime negotiation is a capability AGTP enables, not a behavior it requires.¶
5.4. Connection Model
AGTP uses a persistent session model by default, reflecting the reality that agents typically execute multi-step workflows rather than isolated single requests. An AGTP session is established with a single TLS handshake including agent identity assertion, persists across multiple method exchanges, carries a Session-ID header identifying the agent's task context, and terminates on explicit session close or inactivity timeout (RECOMMENDED minimum: 60 seconds).¶
Per-request (stateless) mode is supported for constrained environments. In stateless mode, agent identity headers MUST be present on every individual request.¶
5.4.1. Wire-Format Framing
AGTP requests and responses MUST be framed by an explicit
Content-Length header. Content-Length is the sole signal of request
and response completion. Receivers MUST treat the message as
complete when, and only when, the declared number of body octets has
been read after the header terminator.¶
AGTP sessions running over TLS MUST NOT use socket-level half-close
(shutdown(SHUT_WR) or equivalent) to signal end-of-request. The TLS
close_notify alert that results from a half-close terminates the
secure session before the peer can transmit a response, producing a
truncation that is indistinguishable at the application layer from a
malicious downgrade. Implementations that require an explicit completion
signal in addition to Content-Length MUST rely on the AGTP
session-close semantics described in Section 7.2.11 or on
transport-level FIN after the full response has been received.¶
Chunked transfer encoding is not used in AGTP. Streaming method
responses are framed by repeated Content-Length-delimited messages
within a single AGTP session.¶
5.5. Header Format
The AGTP wire format puts the protocol version and method on the
request line (AGTP/1.0 METHOD PATH\r\n) and the status code on
the response line (AGTP/1.0 STATUS STATUS-TEXT\r\n). Headers
MUST NOT carry information already present on those lines:
there is no AGTP-Version header, no AGTP-Method header, no
AGTP-Status header. The version and method are read from the
request line; the status is read from the response line.¶
5.5.1. Request Line Grammar
The request line carries three tokens separated by single space characters, terminated by CRLF:¶
request-line = agtp-version SP method SP request-target CRLF
agtp-version = "AGTP/1.0"
method = token ; per {{AGTP-API}} catalog
request-target = path-absolute [ "?" query ]
; per RFC 3986
path-absolute = "/" [ segment-nz *( "/" segment ) ]
; per RFC 3986 Section 3.3
query = *( pchar / "/" / "?" )
; per RFC 3986 Section 3.4
¶
The request-target follows the path-absolute production of
[RFC3986], optionally followed by ? and a query string. Path
and query are parsed as separate tokens: the path is everything
from the leading / up to the first ? or end-of-line; the query
string is everything after the first ? up to the CRLF.¶
Servers MUST parse path and query as separate tokens before
dispatch, matching [RFC3986] URI generic syntax. Implementations
MUST NOT treat a ?-prefixed query as part of the path; the
path-grammar enforcement in [AGTP-API] applies to the path
component only.¶
URI fragments (#anchor) MUST NOT appear on the request line.
A request line containing # MUST be rejected at the wire
layer as malformed.¶
The path-pattern grammar (template parameters in {param} form,
verb-leakage prevention, structural minimums) is normatively
specified in [AGTP-API]. v07 servers MUST implement the
AGTP-API path grammar.¶
5.5.2. Response Line Grammar
The response line carries three tokens separated by single space characters, terminated by CRLF:¶
response-line = agtp-version SP status-code SP status-text CRLF
agtp-version = "AGTP/1.0"
status-code = 3DIGIT ; per {{status-codes}}
status-text = *( VCHAR / SP / HTAB )
; freeform reason phrase
¶
The status-text is informational and MUST NOT be used by
servers or clients for protocol decisions; the numeric status-code
is authoritative.¶
5.5.3. Request Headers
| Field | Required | Description |
|---|---|---|
| Agent-ID | MUST for non-anonymous requests | Canonical 256-bit identifier of the invoking agent. Identifies the request's source, not its target. Servers reject requests lacking Agent-ID against non-anonymous endpoints. |
| Authority-Scope | MAY | Scopes the agent claims for this specific request. When present, every claimed scope MUST be a subset of the scopes declared in the agent's identity document; servers MUST validate and MUST return 262 Authorization Required with body code scope-claim-invalid on failure. When absent, the request inherits the full scope set of the agent's identity document. The server uses the claimed (or inherited) set when evaluating the endpoint's required_scopes. |
| Session-ID | MAY | Opaque session identifier grouping multiple requests into an operational session. The protocol assigns no semantics; servers pass the value through to handlers via the endpoint context. |
| Task-ID | MAY | Opaque identifier tracing a specific task or operation across multiple requests. Useful for audit correlation. Servers SHOULD echo this value in the response Task-ID header. |
| Delegation-Chain | MAY (reserved) | Reserved for delegated-authority scenarios. Format, validation, and chain-of-trust semantics are not specified in this revision and are anticipated in a future revision. v00 servers MAY reject requests carrying this header with 501 Not Implemented (HTTP standard semantics); reject responses SHOULD carry a body indicating the unsupported feature. |
| Merchant-ID | MUST on PURCHASE | Canonical identifier of the intended merchant counterparty. See [AGTP-MERCHANT]. |
| Merchant-Manifest-Fingerprint | MUST on PURCHASE | SHA-256 fingerprint of the Merchant Manifest Document verified by the requesting agent. Receiving server MUST reject with 458 if this does not match its current manifest. See [AGTP-MERCHANT]. |
| Intent-Assertion | SHOULD on PURCHASE | Detached JWT [RFC7519] carrying signed principal-authorized purchase intent. Forwardable to payment networks as standalone evidence. See [AGTP-MERCHANT]. |
| Cart-Digest | MAY | Cryptographic digest of a structured cart returned by a prior QUOTE invocation. Binds a PURCHASE to a previously quoted cart without retransmission of line-item detail. See [AGTP-MERCHANT]. |
5.5.4. Response Headers
| Field | Required | Description |
|---|---|---|
| Server-ID | MUST | Canonical identifier of the server that produced the response. MUST be populated from the server's configured server_id on every response. Useful for audit, load-balanced deployments, and verifying which server processed a request. |
| Task-ID | SHOULD when present in request | Echo of the request's Task-ID header to correlate response with request. Absent when the request did not carry Task-ID. |
| Attribution-Record | SHOULD when manifest signing is configured | JWS-signed attestation of the response's origin. Carries the server ID, response timestamp, request hash (for replay detection), and response status. Coupled with manifest signing: a server with a configured signing key emits Attribution-Record on every response; a server without one omits it. |
| Continuation-Token | MAY | Token for retrieving additional results in streaming contexts. |
| Supported-Methods | SHOULD (on session open) | List of AGTP methods supported by this server. |
| Cost-Estimate | MAY | Estimated resource consumption in Budget-Limit unit format. Returned by QUOTE; MAY appear on any response as an informational signal. |
| Attestation-Evidence | MAY | RATS attestation evidence token or reference URI per [RFC9334]. Format indicated by attestation_type in response body: rats-eat, rats-corim, or rats-uri. |
Implementations MAY emit implementation-specific headers, by
convention prefixed X-, but such headers have no protocol
semantics; agents MUST NOT rely on them and servers MUST
NOT require them.¶
5.5.5. Headers Reserved for Future Revisions
The following headers were specified in earlier drafts and are reserved for future revisions. v00 servers MUST NOT require them and MAY ignore them when received:¶
| Field | Status | Anticipated Use |
|---|---|---|
| Priority | Reserved (v01+) | Request priority hint: critical, normal, background |
| TTL | Reserved (v01+) | Maximum acceptable response latency in milliseconds. Pairs with status code 408 Timeout. |
| Budget-Limit | Reserved (v01+) | Maximum resource budget per invocation. Pairs with status code 456 Budget Exceeded. |
| AGTP-Zone-ID | Reserved (v01+) | Network zone boundary constraint. Pairs with status code 457 Zone Violation; SEP-enforced. |
| Content-Schema | Reserved (v01+) | URI reference to JSON Schema describing the request body structure. |
| Telemetry-Export | Reserved (v01+) | OTLP endpoint URI for metric export, or inline to receive metrics embedded in the response Attribution-Record. |
Status codes 456 Budget Exceeded and 457 Zone Violation remain
allocated in the AGTP Status Code Registry; their normative
trigger headers (Budget-Limit, AGTP-Zone-ID) are deferred to a
future revision. v00 servers MAY enforce equivalent semantics
through implementation-specific mechanisms and return the
corresponding status code, but the wire-level headers are not part
of the v00 contract.¶
5.5.6. Retired Headers
The following headers appeared in earlier drafts and have been removed. Implementations MUST NOT emit them and MAY reject requests that carry them:¶
| Field | Reason for Removal |
|---|---|
| AGTP-Version | Redundant with the request and response lines (which carry AGTP/1.0). |
| AGTP-Method | Redundant with the request line (which carries the method token). |
| AGTP-Status | Redundant with the response line (which carries the status code). |
| Principal-ID | Redundant: the principal is recorded in the agent identity document referenced by Agent-ID. |
| Server-Agent-ID | Renamed to Server-ID. The earlier name conflated server identity with agent identity. |
5.6. Status Codes
AGTP defines its own status code space. Codes 261 through 263, 455 through 465, and 550 through 555 are AGTP-specific with no HTTP equivalent and are registered in the IANA AGTP Status Code Registry (see Section 9.3). AGTP-specific code numbers are deliberately chosen from ranges unassigned in the IANA HTTP Status Code Registry to avoid semantic collision with HTTP status codes that may appear in payloads carried by AGTP method invocations.¶
The AGTP status code model carries four structural rejection codes (404, 405, 459, 460) that together cover the failure surface for contract-level invocation. Each code is independently actionable: 404 indicates the path does not exist on this server; 405 indicates the verb and path are each individually valid but the server does not expose this combination; 459 indicates the verb is not in the AGTP-API approved method catalog; 460 indicates the path violates AGTP-API path grammar. The structural rejection model and the curated method catalog are specified in [AGTP-API].¶
| Code | Name | Meaning |
|---|---|---|
| 200 | OK | Method executed successfully |
| 202 | Accepted | Method accepted; execution is asynchronous |
| 204 | No Content | Method executed; no response body |
| 261 | Negotiation In Progress | PROPOSE evaluation in progress; agent MAY poll for terminal status. See [AGTP-API]. |
| 262 | Authorization Required | Request requires credential establishment, additional authorization scope, or consent that is not yet present. Covers PROPOSE-time authorization, missing scope at endpoint dispatch, ad-hoc method invocation requiring wildcards consent, and discovery requests blocked by anonymous-discovery policy. See [AGTP-API]. |
| 263 | Proposal Approved | PROPOSE accepted; the proposed endpoint has been synthesized and instantiated. Response body carries the synthesized endpoint contract. See [AGTP-API]. |
| 400 | Bad Request | Malformed AGTP request |
| 401 | Unauthorized | Agent-ID not recognized or not authenticated |
| 403 | Forbidden | Agent lacks authority for requested action per Authority-Scope |
| 404 | Not Found | The path does not exist on this server. No endpoints are registered under it. |
| 405 | Method Not Allowed | The method is recognized and the path is valid, but the server's policy or registry does not expose this combination. The response body MUST list allowed methods for the path and any redirects from the manifest's policies.methods sub-block. The agent MAY PROPOSE the combination if it is not exposed by policy. AGTP-specific. See [AGTP-API]. |
| 408 | Timeout | TTL exceeded before method could execute. AGTP-specific semantics; see below. |
| 409 | Conflict | Method conflicts with current state |
| 410 | Gone | Agent has been Revoked or Deprecated; canonical Agent-ID is permanently retired. AGTP-specific semantics; see below. |
| 422 | Unprocessable | Request well-formed but semantically invalid |
| 429 | Rate Limited | Agent is exceeding permitted request frequency |
| 455 | Scope Violation | Requested action is outside declared scope. Generic scope-violation code for cases not covered by the dedicated scope codes (429 rate-limit, 456 budget, 457 zone, 262 authority). Token-based or query-based scope violations are typical applications. AGTP-specific. |
| 456 | Budget Exceeded | Method execution would exceed the Budget-Limit declared in the request. AGTP-specific. |
| 457 | Zone Violation | Request would route outside the AGTP-Zone-ID boundary. SEP-enforced. AGTP-specific. |
| 458 | Counterparty Unverified | PURCHASE counterparty failed merchant identity verification: Merchant-ID absent, Merchant-Manifest-Fingerprint mismatch, or merchant in non-Active lifecycle state. AGTP-specific. See [AGTP-MERCHANT]. |
| 459 | Method Violation | The method name is not in the AGTP-API approved method catalog. The method itself is the problem. AGTP-specific. See [AGTP-API]. |
| 460 | Endpoint Violation | The endpoint path violates AGTP-API path grammar. A path segment matches an approved method name, indicating method-name leakage into the path. AGTP-specific. See [AGTP-API]. |
| 461 | Reserved | Reserved for AGTP expansion. |
| 462 | Reserved | Reserved for AGTP expansion. |
| 463 | Proposal Rejected | The service cannot or will not instantiate the proposed endpoint. Returned in response to PROPOSE. Response body MUST carry a structured reason (e.g., out-of-scope, policy-refused, composition-impossible, ambiguous, synthesis-disabled) and MAY carry an optional counter-proposal. AGTP-specific. See [AGTP-API]. |
| 464 | Reserved | Reserved for AGTP expansion. |
| 465 | Reserved | Reserved for AGTP expansion. |
| 500 | Server Error | Internal failure in the responding system |
| 503 | Unavailable | Responding agent or system temporarily unavailable or Suspended |
| 550 | Delegation Failure | A delegated sub-agent failed to complete the requested action. AGTP-specific. |
| 551 | Authority Chain Broken | Delegation chain contains an unverifiable or broken identity link. AGTP-specific. |
| 552 | Reserved | Reserved for AGTP expansion. |
| 553 | Reserved | Reserved for AGTP expansion. |
| 554 | Reserved | Reserved for AGTP expansion. |
| 555 | Reserved | Reserved for AGTP expansion. |
The four structural rejection codes (404, 405, 459, 460) form the contract-level structural failure surface and have distinct recovery semantics. A 459 tells the agent to choose a different method from the AGTP-API catalog. A 460 tells the agent to restructure the path so that no path segment is a method name. A 405 tells the agent the method and path are each individually valid but the failure is a server policy decision; a different method-and-path combination might work, or PROPOSE might negotiate access. A 404 tells the agent the path does not exist on this server at all. Beyond the structural surface, 463 (Proposal Rejected) is the runtime negotiation rejection: the agent's PROPOSE was structurally valid but the server cannot or will not synthesize the requested endpoint.¶
Status code 262 (Authorization Required) consolidates the
authority-related rejection conditions an agent might encounter
when interacting with a server. It is returned when: (a) a PROPOSE
request requires credential establishment before the server will
evaluate it; (b) an endpoint invocation requires Authority-Scope
that the agent has not declared; (c) an ad-hoc method invocation
requires wildcards consent (wildcards: true on the agent identity
document and wildcards_accepted: true in server policy) that is
absent on either side; or (d) a discovery request is blocked by
server policy that requires authenticated identity for manifest
retrieval. The response body MUST identify which condition
applies so the agent knows what to remediate.¶
Status code 455 (Scope Violation) is a generic scope-violation signal for cases not covered by the dedicated scope codes (429 rate-limit, 456 budget, 457 zone, 262 authority). Typical applications include token-based scope violations and query-based scope violations, where an operator defines a scope dimension outside the standard set. Authority-Scope violations specifically are signaled with 262, not 455. Status code 456 (Budget Exceeded) is the dedicated code for resource-consumption scope violations: the agent's requested action would consume resources beyond what the principal authorized for this invocation. Status code 457 (Zone Violation) is returned by SEPs when a DELEGATE request would route to an agent outside the declared zone boundary. Status code 458 (Counterparty Unverified) is returned on PURCHASE invocations when the receiving server cannot verify that the requesting agent has performed valid merchant identity verification against the server's current Merchant Manifest Document, or when the merchant is in a non-Active lifecycle state; see [AGTP-MERCHANT].¶
Status code 551 (Authority Chain Broken) is returned when a server processing a delegated request cannot reconstruct a valid delegation sequence from the delegating agent to the requesting agent. The specific unverifiable link SHOULD be identified in the response body. Status code 408 (Timeout) is reused from HTTP with AGTP-specific semantics: it signals that the method's declared TTL expired before execution completed, distinct from HTTP's request-timeout semantics. Status code 410 (Gone) is reused from HTTP with AGTP-specific semantics: it indicates that an Agent-ID has been permanently retired through revocation or deprecation of its Agent Genesis, distinct from HTTP's resource-removed semantics. The canonical Agent-ID MUST NOT be retried. All AGTP-specific status codes are operational signals, not protocol errors, and MUST be logged for audit purposes.¶
5.7. Wire Format and Content-Type
AGTP request and response bodies are encoded as JSON or YAML. The following media types are defined by AGTP. Their IANA registration status varies; see the IANA Considerations section for the authoritative status table.¶
| Media Type | Use | IANA Status |
|---|---|---|
application/vnd.agtp+json
|
AGTP method request/response bodies (JSON) | Planned (this document) |
application/vnd.agtp+yaml
|
AGTP method request/response bodies (YAML) | Planned (this document) |
application/vnd.agtp.identity+json
|
Agent Identity Document (JSON) | Vendor-tree registration submitted |
application/vnd.agtp.identity+yaml
|
Agent Identity Document (YAML) | Vendor-tree registration submitted |
Two additional media types are defined in the companion AGTP-API
specification [AGTP-API] and registered concurrently:
application/vnd.agtp.manifest+json (server manifest) and
application/vnd.agtp.endpoint+json (endpoint definition). Both
are listed in the master Media Type Registry section of this
document for cross-document discoverability.¶
Implementations MUST include the appropriate Content-Type on all AGTP requests and responses that carry a message body. Responses with no body (e.g., 204 No Content) MUST NOT include a Content-Type header. Binary or streaming extensions MAY define additional Content-Type values as part of their companion specifications.¶
EXECUTE method invocations carry application-layer payloads whose
Content-Type is set by the carried application protocol (for example,
application/vnd.mcp.tools+json for MCP tool invocations carried over
AGTP). The AGTP server dispatches based on the carried Content-Type; see
Section 7.2.7 and Section 8.¶
The common structure for AGTP method-level request bodies:¶
{
"method": "QUERY",
"task_id": "task-0042",
"session_id": "sess-a1b2c3d4",
"parameters": { },
"context": { }
}
¶
And for AGTP method-level response bodies:¶
{
"status": 200,
"task_id": "task-0042",
"result": { },
"attribution": { }
}
¶
5.8. Early Implementations
AGTP is a proposed specification. No production implementations exist at the time of this writing. The author encourages early prototype implementations to validate the protocol design, identify gaps, and generate feedback prior to IETF working group submission.¶
-
If you are building an AGTP prototype or reference implementation, please share your findings via the feedback channel listed on the cover of this document. A reference implementation in Python and/or Go is planned as open-source software concurrent with or shortly after IETF I-D submission. Implementation reports are welcome and will be incorporated into subsequent draft revisions.¶
Implementers wishing to experiment before final IANA port assignment SHOULD use a locally-chosen port from the Dynamic Ports range (49152-65535) on developer-controlled endpoints. Implementations MUST NOT publish or document any specific port number as an AGTP-associated value until IANA assignment is complete.¶
The ACTIVATE method extension, which binds .nomo governed agent packages
to AGTP as a first-class activation operation, is described in a companion
document and is implemented as an optional extension. Core AGTP
implementations need not support ACTIVATE to be compliant with this
specification.¶
6. Agent Identity, URI Structure, and Registration
6.1. URI Structure and Resolution Mechanics
6.1.1. Foundational Principle
AGTP identity is agent-first. Every agent is identified by a canonical
Agent-ID: a 256-bit cryptographic identifier derived from the agent's
Agent Genesis hash at ACTIVATE time. The canonical Agent-ID is the
authoritative identifier in every AGTP protocol operation. It appears
in the Agent-ID header of every request, is the key in the registry,
and is the cross-layer reference linking the AGTP Agent Certificate
extension to the governance-layer Agent Genesis.¶
All other agent identification forms recognized by AGTP, including domain-anchored URIs and Web3 resolution targets, are aliases that resolve to a canonical Agent-ID. In the event of any conflict between an alias and a canonical Agent-ID, the canonical Agent-ID MUST be treated as authoritative.¶
AGTP URIs are addresses, not filenames. File format suffixes
(.agtp) MUST NOT appear in canonical agtp:// URIs. A URI
resolves to an Agent Identity Document or Agent Namespace Document
derived from the underlying package; it does not expose or serve the
package itself.¶
Implementations MUST treat any URI containing a file extension in the path as non-canonical and SHOULD issue a 301 Moved Permanently redirect to the canonical form prior to resolution.¶
The .agent suffix and the .nomo suffix are file format
designators for AGTP agent packages; they are not URI hostname
labels. Earlier drafts admitted hostname forms ending in .agent
or .nomo as agent-native hierarchical TLDs; this revision removes
that admission. Hostnames in AGTP URIs are standard DNS hostnames
per RFC 3986.¶
6.1.2. Canonical URI Forms
AGTP is identity-first. Agent-IDs are canonical and content-addressed;
hosting is incidental. Form 1 (agtp://[Agent-ID]) is the canonical
identity form. Form 1a (agtp://[Agent-ID]@[host]) provides direct
addressing for cases where the resolver does not yet know how to
reach the canonical ID, and bridges the gap between identity-first
addressing and the existing DNS-based reachability infrastructure.¶
Forms 2 and 2a address servers and organizations rather than specific agents, providing entry points for discovery operations that do not target a named agent. Forms 3 and 4 address agents by local name within a domain's namespace; they differ only in deployment convention.¶
Form 1. Canonical identity (cryptographic, authoritative): agtp://[Agent-ID] Form 1a. Canonical identity with explicit host (direct addressing): agtp://[Agent-ID]@[host] Form 2. Server-level discovery (specific server endpoint): agtp://[host] Form 2a. Organization-level discovery (DNS-registered domain root): agtp://[domain] Form 3. Domain-anchored agent (domain owns the namespace): agtp://[domain]/agents/[agent-name] Form 4. Subdomain-anchored agent (dedicated AGTP subdomain): agtp://agtp.[domain]/agents/[agent-name]¶
The port portion of any AGTP URI is OPTIONAL. When omitted, AGTP clients MUST use the IANA-assigned default port 4480. The port is included in URIs only for non-default deployments and appears in the host component (Form 1a host or Form 2 host); ports MUST NOT appear in Form 2a domain components or in Forms 3/4 domain components.¶
6.1.2.1. Form 1 — Canonical Identity
Form 1 carries only the 64-character lowercase hex representation of the Agent Genesis SHA-256 hash:¶
agtp://7f3a9c2d...e1f8b0a4¶
Form 1 resolves to a signed Agent Identity Document through any verification path declared in the agent's registry record (Section 5.2). The canonical ID is self-describing: any AGTP-aware governance platform, transparency log, or resolution service can return the Identity Document given the canonical ID alone, without prior knowledge of which organization, domain, or blockchain the agent is registered under.¶
6.1.2.2. Form 1a — Canonical Identity with Explicit Host
Form 1a embeds an explicit host alongside the canonical Agent-ID:¶
agtp://7f3a9c2d...e1f8b0a4@agents.acme.com agtp://7f3a9c2d...e1f8b0a4@192.0.2.42 agtp://7f3a9c2d...e1f8b0a4@agents.acme.com:9999¶
The Agent-ID identifies; the host tells the resolver where to reach a server that can return the Identity Document or accept method invocations. Form 1a is the preferred form when:¶
-
The agent has been issued a canonical Agent-ID but is not yet registered with an AGTP discovery service.¶
-
The resolver does not yet have a path from canonical Agent-ID to endpoint (no DNS anchor, no transparency log entry, no governance-platform resolution).¶
-
A client wishes to address an agent by ID directly without round- tripping through a registry.¶
Form 1a does not weaken the identity-first model. The canonical
Agent-ID remains authoritative; the host portion is treated as a
resolution hint and MUST NOT be used to derive identity. If the
host returns an Identity Document whose agent_id field does not
match the canonical Agent-ID in the URI, the response MUST be
rejected.¶
6.1.2.3. Form 2 — Server-Level Discovery
Form 2 addresses a specific server endpoint without naming an agent:¶
agtp://agents.acme.com agtp://agtp-edge-7.us-east.acme.com agtp://192.0.2.42 agtp://[2001:db8::42] agtp://agents.acme.com:9999¶
The host component MAY be any RFC 3986 host: a registered hostname, an FQDN, an IPv4 address, an IPv6 address, optionally followed by a port. Form 2 is used for diagnostic operations addressing a specific server instance (a particular edge node, a specific deployment in a load-balanced fleet, a server identified by IP) and for method invocations that target server-level state rather than agent-level state.¶
6.1.2.4. Form 2a — Organization-Level Discovery
Form 2a addresses an organization's AGTP presence via its registered DNS domain:¶
agtp://acme.com agtp://example.org¶
The domain component MUST be a DNS-registered domain name with at least one label. IP addresses, ports, and userinfo components MUST NOT appear in Form 2a; URIs carrying any of these are Form 2, not Form 2a.¶
Form 2a resolution proceeds via standard DNS lookup of the domain, then AGTP connection establishment on port 4480 against the resolved address. The response is the organization's canonical server manifest. Whether the domain resolves to a single AGTP server, a load-balanced fleet, or a CDN-fronted edge is a deployment concern; Form 2a does not commit the caller to any specific topology.¶
The syntactic distinction between Form 2 (any RFC 3986 host, possibly with port) and Form 2a (bare DNS domain) corresponds to a semantic distinction. Form 2 addresses a specific reachable server; Form 2a addresses whichever server the organization's DNS currently points at. Both are valid; they differ in caller intent.¶
6.1.2.5. Forms 3 and 4 — Domain-Anchored Agents
Forms 3 and 4 address an agent by local name within a domain's agent namespace:¶
Form 3: agtp://acme.com/agents/bookbot Form 4: agtp://agtp.acme.com/agents/bookbot¶
Form 3 places the agent under the organization's primary domain.
Form 4 places the agent under a dedicated agtp. subdomain.
Resolution semantics are identical: the AGTP server at the domain
(or subdomain) consults its hosted_agents manifest entries and
returns the canonical Agent-ID for the named local agent. The local
agent name is unique within the domain's namespace; the same name
under a different domain is a different agent.¶
Forms 3 and 4 differ only in deployment convention. Operators choose based on their infrastructure preferences: organizations that prefer to keep agent-protocol traffic on a dedicated subdomain use Form 4; organizations that operate AGTP as the canonical face of their primary domain use Form 3. The protocol treats them as equivalent.¶
In both forms the domain component MUST be a DNS-registered domain name; IPs and ports MUST NOT appear. Direct addressing with explicit host is available via Form 1a if the resolver knows the canonical Agent-ID.¶
6.1.3. URI Grammar
The URI grammar is defined in ABNF as:¶
AGTP-URI = "agtp://" agtp-locator
agtp-locator = agent-id [ "@" host ] ; Form 1, 1a
/ host ; Form 2
/ domain ; Form 2a
/ domain "/agents/" agent-name ; Form 3
/ "agtp." domain "/agents/" agent-name ; Form 4
agent-id = 64HEXDIG ; lowercase hex
host = IP-literal / IPv4address / reg-name [ ":" port ]
; per RFC 3986
domain = label *("." label) ; DNS-registered, no port
label = ALPHA *( ALPHA / DIGIT / "-" )
agent-name = 1*( ALPHA / DIGIT / "-" / "_" )
port = 1*DIGIT
; OPTIONAL; defaults to 4480
¶
AGTP URIs are addresses, not filenames. File extensions in the
path MUST NOT appear in canonical agtp:// URIs. A URI
resolves to an Agent Identity Document, a server manifest, or an
endpoint response derived from server state; it does not expose or
serve a package file.¶
Implementations MUST treat any URI containing a file extension in the path as non-canonical and SHOULD issue a 301 Moved Permanently redirect to the canonical form prior to resolution.¶
6.1.4. Method-on-URI Invocation Pattern
AGTP method invocations target a URI plus an optional endpoint path. The conceptual pattern is:¶
METHOD agtp://[locator]/[endpoint-path]¶
The locator addresses an agent (Forms 1, 1a, 3, 4), a server (Form 2), or an organization (Form 2a). The endpoint path is the path portion of the request line as specified in Section 5.7. For server-level and organization-level discovery operations the endpoint path is omitted (target-less DISCOVER); for agent-level operations the endpoint path identifies the endpoint to invoke.¶
The wire-level encoding of this pattern is the AGTP request line (method and path) plus the connection target (host derived from the URI). The URI is the agent-facing addressing notation; the wire format is what travels over TLS.¶
6.1.5. Web3 Verification Anchors
AGTP supports Web3-anchored verification paths for canonical Agent-IDs whose underlying Agent Genesis is registered through a blockchain-based verification service. The verification path is declared in the agent's registry record per Section 6.1.9 and does not affect URI syntax: a Web3-anchored agent is addressed by its canonical Agent-ID via Form 1 or Form 1a like any other agent.¶
Integration with specific Web3 naming and resolution systems is specified in [AGTP-WEB3].¶
6.1.6. Non-Canonical Forms and Redirect Behavior
The following non-canonical forms SHOULD be redirected to their canonical equivalents. Implementations MUST NOT serve package contents in response to any URI form.¶
| Received URI | Canonical Redirect Target |
|---|---|
agtp://acme.tld/agents/customer-service.agent
|
agtp://acme.tld/agents/customer-service
|
agtp://acme.tld/agents/customer-service.nomo
|
agtp://acme.tld/agents/customer-service
|
agtp://acme.tld/agents/customer-service.agtp
|
agtp://acme.tld/agents/customer-service
|
6.1.7. Query Parameters for Format Selection
All AGTP URI resolution requests accept an optional format query
parameter controlling the serialization of the returned document.¶
| Query Parameter | Returned Representation |
|---|---|
| (none) | Agent Identity Document, human-readable application/vnd.agtp+json
|
?format=manifest
|
Agent Identity Document, human-readable application/vnd.agtp+json
|
?format=json
|
Agent Identity Document, compact application/vnd.agtp+json
|
?format=certificate
|
Agent Genesis fields only, application/vnd.agtp+json
|
?format=status
|
Lifecycle state and operational status only, application/vnd.agtp+json
|
All format variants return signed application/vnd.agtp+json content.
The ?format=json parameter is intended for programmatic consumers.
The default returns the full human-readable manifest suitable for
browser rendering by an AGTP-aware client.¶
6.1.8. Resolution Mechanics
AGTP URI resolution proceeds according to the URI form presented.
Form 1 and Form 1a (canonical agent identity) resolve through a
governance-platform registry or transparency-log lookup. Forms 2
and 2a (server and organization discovery) resolve through
direct AGTP connection establishment to the host or domain. Forms
3 and 4 (domain-anchored agents) resolve through DNS to an AGTP
server whose hosted_agents manifest maps the URI's local agent
name to a canonical Agent-ID.¶
6.1.8.1. Form 1 Resolution (Canonical ID)
When an AGTP resolver receives a URI of the form
agtp://[256-bit-hex-id], it MUST perform the following steps:¶
-
Parse and validate the canonical Agent-ID. If the identifier is malformed (length, character set), return
400 Bad Requestwith error codeinvalid-canonical-id.¶ -
Query the agent's governance platform registry for the record associated with the canonical Agent-ID. If the resolver does not know which governance platform holds the record, it MAY query a transparency log per Section 5.2 to locate the record.¶
-
Verify the registry record lifecycle state. If Suspended, return
503 Service Unavailablewith lifecycle state in the response body. If Revoked or Deprecated, return410 Gonewith lifecycle state and revocation timestamp.¶ -
Retrieve the agent's package (
.agentor.nomo) from the package store referenced by the registry record.¶ -
Verify the package integrity hash before proceeding. If integrity verification fails, return
500 Internal Errorwith error codepackage-integrity-failure. MUST be logged.¶ -
Extract the embedded manifest from the verified package.¶
-
Sign the Identity Document using the governance platform's signing key. Return the signed
application/vnd.agtp+jsondocument in the format specified by the query parameter.¶
Form 1 resolution does not require prior knowledge of an organization domain, a DNS record, or a Web3 naming anchor. The canonical Agent-ID is sufficient input.¶
6.1.8.2. Forms 2 and 2a Resolution (Server and Organization Discovery)
When an AGTP resolver receives a URI of Form 2 (agtp://[host]) or
Form 2a (agtp://[domain]), it MUST perform the following
steps:¶
-
Parse and validate the URI. Form 2 admits any RFC 3986 host with optional port. Form 2a admits a DNS-registered domain name with at least one label and MUST NOT include port or userinfo components.¶
-
Open an AGTP/TLS connection to the host or domain. For Form 2a, resolve the domain via DNS first. For both forms, AGTP traffic targets port 4480 unless an explicit non-default port appears in the URI (Form 2 only).¶
-
Issue the agent's intended method (typically
DISCOVERfor discovery operations) with no agent-identifying target.¶ -
The server returns its server manifest (for Form 2 or Form 2a addressed against a server's primary endpoint) or whatever server-level response the invoked method produces.¶
Forms 2 and 2a do not resolve to a canonical Agent-ID; they target server or organization state directly. They do not participate in the canonical-Agent-ID identity model.¶
6.1.8.3. Forms 3 and 4 Resolution (Domain-Anchored Agents)
When an AGTP resolver receives a URI of Form 3 or Form 4, it MUST perform the following steps:¶
-
Parse and validate the URI. If the URI is malformed, return
400 Bad Requestwith error codeinvalid-uri-form.¶ -
Resolve the domain (Form 3) or
agtp.[domain]subdomain (Form 4) via DNS, then open an AGTP/TLS connection to port 4480.¶ -
Look up the
agent-namefrom the URI path against the server'shosted_agentsmanifest entries to obtain the canonical Agent-ID. If no matching agent is found, return404 Not Found.¶ -
Continue with Form 1 resolution steps 3 through 7 using the resolved canonical Agent-ID.¶
The package's executable content, code, logic, and any fields not included in the Identity Document schema MUST NOT be returned at any step of any resolution path. URI resolution exposes identity and status exclusively.¶
6.1.9. Trust Tiers and Verification Paths
The Agent Genesis carries a trust_tier field (one of 1
Verified, 2 Org-Asserted, 3 Experimental) and a
verification_path field (one of dns-anchored, log-anchored,
hybrid, or org-asserted) recording how the agent's identity
was anchored at ACTIVATE time. These fields are surfaced in the
Agent Identity Document and consulted by Scope-Enforcement Points,
governance gateways, and peer agents during runtime authority
decisions.¶
Tier 1 agents are eligible for the full Authority-Scope vocabulary,
delegation chains, financial transactions, and multi-organization
collaboration. Tier 1 verification requires exactly one of three
verification paths (dns-anchored, log-anchored, hybrid) to
succeed at ACTIVATE time, each backed by distinct evidence (DNS
TXT record, transparency log inclusion proof, or DNS + blockchain
signature combination). All Tier 1 paths produce identity
attestations of equivalent strength.¶
Tier 2 agents declare organizational affiliation without
cryptographic verification (verification_path: org-asserted) and
carry the trust_warning: "verification-incomplete" signal. Tier
3 agents are experimental and confined to development environments.¶
The normative verification procedures, evidence requirements, tier assignment rules, and Authority-Scope eligibility constraints are specified in [AGTP-TRUST].¶
6.1.10. Subdomain Deployment Pattern
Organizations SHOULD deploy AGTP endpoints at a dedicated
subdomain following the pattern agtp.[organization-domain.tld]
(e.g., agtp.acme.tld). This is the recommended enterprise
deployment pattern: it provides clean separation between web and
agent infrastructure, allows independent certificate management for
the AGTP endpoint, and is consistent with service-specific subdomain
conventions. An organization with an AGTP subdomain SHOULD also
configure their primary domain to redirect AGTP requests:¶
agtp://acme.tld/agents/customer-service → 301 → agtp://agtp.acme.tld/agents/customer-service¶
6.1.11. The /agents/ Reserved Path Prefix
The path prefix /agents/ is reserved in all agtp:// URIs for
agent namespace operations. Implementations MUST support this
prefix. The registry root at /agents (no trailing label) resolves
to the Agent Namespace Document (see Section 5.4).¶
6.1.12. Collision Prevention
The canonical Agent-ID is the collision-prevention primitive. Two canonical Agent-IDs are distinct if and only if the 256-bit identifiers differ, and the governance platform enforces uniqueness at issuance time by deriving the ID from the Agent Genesis hash.¶
For alias forms, collision prevention operates at the namespace
level. agtp://acme.com/agents/customer-service and
agtp://chrishood.com/agents/customer-service resolve to distinct
canonical Agent-IDs because they are registered under different
DNS domains. Within a single governance zone, the governance
platform enforces uniqueness of agent local names at registration
time.¶
Infrastructure MUST use the canonical Agent-ID for all routing,
logging, and attribution operations. Alias URIs are a display and
discovery layer only. An alias that resolves to a canonical Agent-ID
different from the one carried in the Agent-ID header on a request
MUST cause the request to be rejected with 401 Unauthorized
and MUST be logged.¶
6.1.13. IANA Considerations for the agtp:// URI Scheme
The agtp:// URI scheme is registered with IANA per [RFC7595].
Registration template:¶
- URI scheme name:
-
agtp¶
- Status:
-
Permanent¶
- URI scheme syntax:
-
agtp://[Agent-ID](canonical identity, Form 1)agtp://[Agent-ID]@[host](canonical identity with explicit host, Form 1a)agtp://[host](server-level discovery, Form 2)agtp://[domain](organization-level discovery, Form 2a)agtp://[domain]/agents/[agent-name](domain-anchored agent, Form 3)agtp://agtp.[domain]/agents/[agent-name](subdomain-anchored agent, Form 4)¶ - URI scheme semantics:
-
Identifies an AI agent, an AGTP server, or an organizational AGTP presence operating over the Agent Transfer Protocol. The authoritative agent identification form (Form 1) uses a 256-bit hex-encoded cryptographic identifier derived from the agent's Agent Genesis. Form 1a augments the canonical ID with an explicit host to enable direct addressing in deployments where the canonical ID has not yet been resolved through a registry. Forms 2 and 2a address servers and organizations respectively for discovery operations that do not target a specific agent. Forms 3 and 4 address agents by local name within a domain's namespace and resolve to canonical Agent-IDs through the server's
hosted_agentsmanifest. Forms 3 and 4 differ only in deployment convention.¶ - Applications/protocols that use this URI scheme:
-
Agent Transfer Protocol (this document)¶
- Interoperability considerations:
-
The canonical Agent-ID form is the authoritative agent identity representation. Form 1a embeds a resolution hint alongside the canonical Agent-ID and MUST NOT be used to derive identity. Forms 3 and 4 resolve through DNS to an AGTP server endpoint whose
hosted_agentsmanifest maps the URI path component to a canonical Agent-ID. Forms 2 and 2a do not resolve to canonical Agent-IDs; they target server or organization state directly. Implementations MUST accept canonical Agent-IDs (Form 1) and SHOULD support at least one of the domain-anchored agent forms (Form 3 or Form 4). The port portion of any AGTP URI is OPTIONAL and defaults to the IANA-assigned port 4480 when omitted; ports MAY appear in Form 1a and Form 2 only. File format suffixes (.agtp,.agent,.nomo) MUST NOT appear inagtp://URIs.¶ - Contact:
-
Chris Hood, chris@nomotic.ai¶
- References:
-
This document¶
The agtp:// URI scheme registration is open and unencumbered. No
intellectual property claims apply to the URI scheme itself.¶
6.1.14. .well-known Bootstrap
Organizations operating an AGTP namespace under a DNS domain
SHOULD publish a bootstrap document at the well-known URI
https://{domain}/.well-known/agtp per [RFC8615]. The document is
returned over HTTPS and declares the organization's AGTP presence,
the endpoint at which AGTP traffic should be directed, and any
verification anchors the organization wishes to surface for
discovery.¶
The bootstrap document is JSON with the following fields:¶
{
"agtp_version": "0.7",
"endpoint": "agtp://agtp.acme.tld/agents",
"namespace_root": "agtp://acme.tld/agents",
"verification": {
"dns_anchor": "_agtp.acme.tld",
"log_anchor": null
},
"issuer": "https://ca.acme.tld",
"discovery": {
"namespace_document": "https://agtp.acme.tld/agents.json",
"well_known_updated_at": "2026-04-30T00:00:00Z"
}
}
¶
Resolvers encountering an agtp:// URI for a domain they have not
previously interacted with SHOULD retrieve the bootstrap document
to learn the canonical AGTP endpoint and verification anchors before
attempting AGTP traffic. The bootstrap is advisory; absence of a
.well-known/agtp document does not indicate that the domain has no
AGTP presence, only that the domain has not chosen to advertise one
through this mechanism.¶
A .well-known/agtp document returning HTTP 404 is a positive
declaration that the domain operates no AGTP namespace. Resolvers
MAY cache this negative result for the period indicated by the
HTTP response's Cache-Control directive, with a default of 24
hours when no directive is present.¶
6.2. Agent Namespace Document
6.2.1. Purpose and Scope
The Agent Namespace Document is the index of all Active agents
registered under an organization's governance zone. It is returned
in response to a request targeting the /agents path:¶
agtp://acme.tld/agents agtp://agtp.acme.tld/agents¶
The Agent Namespace Document is not a manually editable file. It is generated and cryptographically signed by the governance platform each time the registry changes. Any Namespace Document that fails signature verification MUST be rejected by the requesting party.¶
6.2.2. Document Schema
{
"document_type": "agtp-namespace",
"schema_version": "1.0",
"org_domain": "acme.tld",
"governance_zone": "zone:acme-internal",
"generated_at": "2026-03-20T14:00:00Z",
"signature": {
"algorithm": "ES256",
"key_id": "agtp-gov-key-acme-01",
"value": "[base64-encoded-signature]"
},
"agents": [
{
"agent_label": "customer-service",
"canonical_id": "3a9f2c1d8b7e4a6f...",
"lifecycle_state": "Active",
"trust_tier": 1,
"cert_status": "Active",
"manifest_uri": "agtp://agtp.acme.tld/agents/customer-service",
"activated_at": "2026-01-15T09:00:00Z",
"last_updated": "2026-03-01T11:30:00Z"
}
],
"total_active": 1,
"namespace_cert_fingerprint": "b2c4d6e8..."
}
The agents array MUST include only agents in Active lifecycle
state. Suspended, Revoked, and Deprecated agents MUST NOT appear
in the Namespace Document.¶
6.2.3. Integrity and Freshness
The Namespace Document MUST include a generated_at timestamp.
Implementations SHOULD treat Namespace Documents older than a
configurable freshness threshold (default: 300 seconds) as stale and
re-request. The governance platform MUST re-sign the Namespace
Document within 60 seconds of any registry change.¶
The signature covers the entire document including generated_at.
Replaying an older signed Namespace Document to conceal a revocation
event is a known attack vector; implementations MUST reject
Namespace Documents with a generated_at timestamp older than the
freshness threshold.¶
6.3. Agent Identity Document and the .agtp Format
6.3.1. Purpose and Scope
The Agent Identity Document is the protocol's canonical representation
of a specific agent's identity, status, and behavioral scope. Prior
versions of this specification referred to this artifact as the Agent
Manifest Document; v07 renames it to the Agent Identity Document to
reflect its IANA-registered media type
(application/vnd.agtp.identity+json) and to clarify its role as the
identity primitive of the protocol. The Identity Document is returned
in response to any AGTP URI resolution request targeting a specific
agent:¶
agtp://[canonical-agent-id] agtp://acme.tld/agents/customer-service agtp://acme.tld/agents/customer-service?format=json¶
The Identity Document is derived from the embedded manifest inside the
agent's .agent or .nomo package. It is not a separate file that can
be independently modified. The governance platform MUST verify the
package integrity hash before extracting and serving the Identity
Document.¶
6.3.2. The Three Document Formats and Their Relationship
.agent ──── open packaging format (patent pending, Chris Hood)
│ Contains: manifest + integrity hash + behavioral
│ trust score + executable content
│
.nomo ──── governed packaging format (patent pending, Chris Hood)
│ Contains: .agent contents + CA-signed cert chain
│ + governance zone binding
│
Both produce ↓ when queried via agtp:// URI
│
.agtp ──── wire-level Identity Document (this spec)
Contains: identity fields, lifecycle state, trust
tier, trust score, behavioral scope, methods,
capabilities, verification anchors
Open format defined by AGTP specification
Content-Type: application/vnd.agtp.identity+json
The .agtp format is the protocol's own document type. It is what
gets transmitted over the wire. The .agent and .nomo formats are
what get deployed. An AGTP implementation is not required to understand
.agent or .nomo packaging internals; it is only required to
produce and consume .agtp Identity Documents.¶
Additional packaging formats MAY be defined by third parties,
provided they can produce conformant .agtp Identity Documents when
queried. Such formats MUST implement the integrity verification
requirement: the Identity Document MUST be derived from a verified
package, not from an independently stored or editable file.¶
6.3.3. Agent Identity Document Schema
The Agent Identity Document carries three categories of fields: REQUIRED fields that every Identity Document MUST contain; RECOMMENDED fields that SHOULD be present when the corresponding information is available; and CONDITIONAL fields that MUST be present when the conditions described later in this section are met.¶
The following fields are REQUIRED in all Agent Identity Documents:¶
{
"agtp_version": "0.7",
"document_type": "agtp-identity",
"document_version": "1.0",
"agent_id": "3a9f2c1d8b7e4a6f0c2d5e9b1a3f7c0d...",
"name": "customer-service",
"description": "Handles customer service requests.",
"principal": "Acme Corporation",
"principal_id": "acme.tld",
"issuer": "https://ca.acme.tld",
"issued_at": "2026-01-15T09:00:00Z",
"updated_at": "2026-03-01T11:30:00Z",
"status": "Active",
"methods": [
"QUERY", "DESCRIBE", "SUMMARIZE",
"EXECUTE", "ESCALATE", "CONFIRM", "NOTIFY"
],
"capabilities": [
"customer-service:tier-1",
"knowledge-base:read",
"ticket-system:create"
],
"scopes_accepted": [
"documents:query",
"knowledge:query",
"calendar:book",
"escalation:route"
],
"trust_score": 0.94
}
Field semantics:¶
-
agtp_version: -
The version of the AGTP protocol the agent speaks. SemVer string.¶
-
document_type: -
Self-identification of the document. MUST be
agtp-identityfor documents conforming to this specification.¶ -
document_version: -
The schema version of the Identity Document itself. SemVer string. Distinct from
agtp_version. Permits the document schema to evolve independently of the protocol.¶ -
agent_id: -
The 256-bit canonical Agent-ID, hex-encoded, derived from the agent's Agent Genesis hash. Authoritative in every AGTP protocol operation.¶
-
name: -
A human-readable label for the agent. Not unique across the AGTP ecosystem; uniqueness is provided by
agent_id.¶ -
description: -
Short prose describing what the agent does.¶
-
principal: -
The human-readable name of the organization or individual that governs the agent.¶
-
principal_id: -
A stable, machine-readable identifier for the principal (typically the principal's primary DNS domain, or a canonical organization identifier).¶
-
issuer: -
The URL of the issuer that signed the Identity Document. The issuer's public key is published at a well-known location under this URL; signature verification details are specified in [AGTP-CERT].¶
-
issued_at: -
ISO 8601 timestamp at which the Identity Document was first issued (typically the moment of ACTIVATE).¶
-
updated_at: -
ISO 8601 timestamp at which the Identity Document was last updated. Promotes freshness checks. MUST be greater than or equal to
issued_at.¶ -
status: -
The lifecycle state of the agent. One of:
Active,Suspended,Revoked,Deprecated,Pending.¶ -
methods: -
The set of AGTP methods the agent's server speaks. MUST include every method the server is prepared to accept. The protocol-defined twelve-method floor (see Section 7.1) MUST be represented if the server claims AGTP conformance.¶
-
capabilities: -
Higher-level service descriptors of what the agent can do. Distinct from
methods: methods are the protocol verbs the server speaks; capabilities are the application-level competencies the agent exposes through those methods.¶ -
scopes_accepted: -
The Authority-Scope tokens the agent will accept on inbound requests. Format and semantics defined in Appendix B.¶
-
trust_score: -
A scalar value on the closed interval [0.0, 1.0] expressing behavioral trust assessed by the agent's principal or governance platform. Computation methodology, freshness requirements, and signature binding are specified in [AGTP-TRUST]. The base AGTP specification carries the field; AGTP-TRUST specifies how the value is produced and verified.¶
The following fields are RECOMMENDED:¶
{
"trust_tier": 1,
"org_domain": "acme.tld",
"governance_zone": "zone:acme-internal",
"verification_path": "dns-anchored",
"cert_fingerprint": "b2c4d6e8...",
"cert_status": "Active",
"dns_anchor_record": "_agtp.acme.tld TXT agtp-zone=...",
"log_inclusion_proof": null,
"package_format": "nomo",
"package_integrity_hash": "sha256:[hash]",
"audit_log_uri": "agtp://agtp.acme.tld/audit/customer-service",
"escalation_policy": "route-to-human-on-scope-limit",
"delegation_permitted": false,
"max_delegation_depth": 0
}
-
trust_tier: -
Tier classification per [AGTP-TRUST]. One of
1,2, or3.¶ -
org_domain: -
The DNS domain under which the agent is registered, when applicable.¶
-
governance_zone: -
The governance zone identifier for the agent's deployment context.¶
-
verification_path: -
The Tier 1 verification path used at ACTIVATE time. One of
dns-anchored,log-anchored, orhybrid. MUST match theverification_pathfield of the underlying Agent Genesis. Detailed semantics specified in [AGTP-CERT].¶ -
cert_fingerprint,cert_status: -
dns_anchor_record: -
Populated when
verification_pathisdns-anchoredorhybrid.¶ -
log_inclusion_proof: -
Populated when
verification_pathislog-anchored. Carries the transparency log inclusion proof per [RFC9162] or the COSE_Sign1 receipt per [RFC9943]. Detailed semantics specified in [AGTP-LOG].¶ -
package_format,package_integrity_hash: -
Deployment-layer metadata identifying the package format and the hash that the Identity Document was extracted from.¶
-
audit_log_uri: -
AGTP URI of the agent's audit log resource.¶
-
escalation_policy,delegation_permitted,max_delegation_depth: -
Policy fields governing the agent's escalation and delegation behavior.¶
The following fields are CONDITIONAL and MUST be present when the conditions described below are met:¶
-
trust_warning: -
REQUIRED when
trust_tieris2, or whentrust_scoreis below the threshold configured by the issuer. Value is a short string identifying the warning class (e.g.,verification-incomplete,trust-score-degraded).¶ -
trust_explanation: -
REQUIRED when
trust_warningis present. Value is a prose description suitable for surfacing to a human operator or in audit logs.¶
The Identity Document is signed by the issuer. v07 carries the signature as a forward reference to [AGTP-CERT], which specifies the signature envelope format (JWS or COSE_Sign1) wrapping the Identity Document. Implementations conforming to this specification MUST verify the signature on the Identity Document against the issuer's published public key before treating any field in the document as authoritative.¶
6.3.4. What the Identity Document Exposes and Does Not Expose
The Agent Identity Document MUST expose:¶
-
The agent's identity (canonical Agent-ID, name, principal, governance zone)¶
-
The agent's current operational status¶
-
The agent's authority scopes accepted¶
-
The agent's supported method vocabulary¶
-
The agent's capability descriptors¶
-
The agent's trust score and trust tier¶
-
The agent's verification anchors and certificate metadata¶
-
The agent's freshness metadata (
issued_at,updated_at)¶
The Agent Identity Document MUST NOT expose:¶
-
Executable code, scripts, or logic¶
-
Model weights or configurations¶
-
Internal API keys or credentials¶
-
Specific authority scope grant tokens¶
-
Session history or prior action logs¶
No AGTP URI resolution path, including any query parameter combination, MUST return package contents beyond the Identity Document schema defined in this section.¶
6.3.5. Identity Document Tamper-Proofing
The tamper-proof guarantee rests on two mechanisms:¶
-
Package integrity hash: Any modification to the package or its embedded manifest invalidates the hash. The governance platform MUST verify this hash before extracting the Identity Document.¶
-
Issuer signature envelope: The issuer signs the extracted Identity Document inside the envelope format defined in [AGTP-CERT]. The signature covers the full document including the
package_integrity_hashfield (when present), creating a verifiable chain from the served document back to the registered package.¶
An Identity Document that fails either verification step MUST be rejected, MUST NOT be served, and the failure MUST be logged.¶
6.4. Browser and Human-Facing Interaction Model
6.4.1. The Separation of Discovery and Execution
agtp:// URI in a browser
│
└─→ Returns Agent Identity Document
Human-readable view of identity and status
Read-only. No execution. No code exposed.
agtp:// session initiated by an agent or AGTP client
│
└─→ Establishes authenticated AGTP session
Method invocations (QUERY, EXECUTE, ESCALATE, etc.)
Full protocol operation - not visible to browsers
The analogy to existing protocol conventions is direct. A mailto:
URI surfaces an address and hands off to a mail client; SMTP carries
the actual messages. Similarly, an agtp:// URI surfaces identity
and status; AGTP carries agent traffic. Browsers do not become AGTP
clients by following an agtp:// link.¶
6.4.2. Browser Behavior for agtp:// URIs
Browsers that encounter an agtp:// URI SHOULD behave as follows:¶
-
If a registered AGTP client is present (OS protocol handler), hand off the URI to that client.¶
-
If the browser supports
agtp://natively or via extension, render the returned Agent Identity Document as a structured human-readable page. The rendered view MUST surface the trust tier indicator prominently, following the visual convention established for TLS trust in the browser chrome.¶ -
If neither condition applies, the browser MAY fall back to a gateway that translates between
https://andagtp://. The gateway MUST preserve all signature and trust tier fields.¶
6.4.3. Human-Readable Identity Document View
When an Agent Identity Document is rendered for human consumption, the following fields MUST be prominently displayed:¶
6.4.4. AGTP Status Sub-Resource
Implementations SHOULD support a status sub-path:¶
agtp://acme.tld/agents/customer-service/status¶
{
"document_type": "agtp-status",
"canonical_id": "3a9f2c1d8b7e4a6f...",
"agent_label": "customer-service",
"org_domain": "acme.tld",
"lifecycle_state": "Active",
"cert_status": "Active",
"last_action_method": "QUERY",
"last_action_timestamp": "2026-03-20T13:58:22Z",
"active_session_count": 3,
"pending_escalations": 0,
"generated_at": "2026-03-20T14:00:00Z"
}
The active_session_count field SHOULD only be included if the
requester has appropriate observability permissions for the governance
zone.¶
6.5. Agent Registration Process
6.5.1. Overview
An agent cannot participate in AGTP until it has been issued an Agent Agent Genesis by a governance platform and assigned a canonical Agent-ID derived from that certificate. Canonical Agent-IDs are issued through the ACTIVATE transaction; they are never self-declared.¶
The Agent Genesis is the genesis record of an agent's legal existence within the AGTP ecosystem. Its relationship to the canonical Agent-ID is analogous to the relationship between a government-issued birth registration and a social security number: the birth event produces a permanent, authoritative identity record, and a durable identifier is derived from it. The identifier follows the agent for its entire lifecycle, including after revocation. It is never reissued to another agent.¶
Any AGTP infrastructure component MUST reject requests carrying an Agent-ID that does not resolve to an Agent Genesis record in an Active lifecycle state in a reachable registry.¶
6.5.2. Agent Genesis Contents
The Agent Genesis is issued by the governance platform at ACTIVATE time and contains the following fields:¶
| Field | Required | Description |
|---|---|---|
agent_id
|
MUST | Unique identifier for the agent |
owner
|
MUST | Human or team responsible for this agent |
archetype
|
MUST | Behavioral category (see archetypes below) |
governance_zone
|
MUST | Environment context (development, staging, production) |
scope
|
MUST | Authorized action types |
issued_at
|
MUST | Timestamp of issuance |
certificate_hash
|
MUST | Cryptographic fingerprint - basis for canonical Agent-ID |
signature
|
MUST | Signed with the org's governance key |
package_ref
|
SHOULD | Reference to the .agent or .nomo package |
trust_tier
|
MUST | Registration tier (1, 2, or 3) |
verification_path
|
MUST (Tier 1) | Path used to verify identity: dns-anchored, log-anchored, hybrid, or org-asserted
|
org_domain
|
SHOULD | DNS-verified or asserted org domain (required for dns-anchored and hybrid) |
org_label
|
SHOULD | Agent-native org label (required for Form 2 hierarchical resolution) |
log_inclusion_proof
|
MUST (log-anchored) | Transparency log inclusion proof (RFC 9162 / RFC 9943) |
6.5.3. Agent Archetypes
The archetype field classifies the agent's behavioral category.
Archetypes inform scope enforcement and observability tooling; an
executor archetype agent exhibiting read-only query patterns, or a
monitor archetype agent attempting booking operations, are anomaly
signals. The archetype field does not restrict scope enforcement,
Authority-Scope headers govern actual permissions at the protocol
level. Archetypes are a classification and observability signal,
not a security boundary.¶
| Archetype | Description | Typical Scope |
|---|---|---|
assistant
|
Conversational agent, read-heavy |
documents:query, knowledge:query
|
analyst
|
Data analysis, read and aggregate |
data:read, data:aggregate
|
executor
|
Takes real-world actions, write-heavy |
booking:*, payments:confirm
|
orchestrator
|
Manages other agents |
delegation:*, agents:*
|
monitor
|
Observational only |
telemetry:read, logs:read
|
6.5.4. Agent Genesis to AGTP Protocol Mapping
Agent Genesis fields map to AGTP protocol elements that surface during request processing. This mapping is the mechanism by which static identity (the Agent Genesis) becomes runtime identity (the protocol session):¶
| Agent Genesis Field | AGTP Protocol Element |
|---|---|
agent_id
|
Agent-ID header on every request |
owner
|
Principal identifier in the agent identity document; not transmitted as a separate header |
scope
|
Full Authority-Scope set in the agent identity document; optionally narrowed per-request via the Authority-Scope header |
certificate_hash
|
Basis for canonical Agent-ID |
The canonical Agent-ID is derived from the certificate_hash. This
chain, package integrity hash → certificate hash → canonical Agent-ID,
ensures that the identifier carried in the Agent-ID header on
every AGTP request is traceable back to the original Agent Genesis
and the human principal who authorized the agent's creation.¶
6.5.5. Registration Tiers
Registration produces an Agent Genesis at one of three trust tiers
(Tier 1 Verified, Tier 2 Org-Asserted, Tier 3 Experimental). Tier 1
registration requires one of three verification paths
(dns-anchored, log-anchored, hybrid) to succeed at ACTIVATE
time, each backed by distinct cryptographic evidence. Tier 2
registration declares organizational affiliation without
cryptographic proof; the resulting Agent Genesis carries
trust_warning: "verification-incomplete". Tier 3 registration is
confined to development environments and produces a locally-scoped
Agent Genesis.¶
The complete tier-specific packaging, evidence, and validation requirements are specified in [AGTP-TRUST].¶
6.5.6. Registration Lifecycle
1. PACKAGE
Author creates .agent or .nomo package containing:
- Embedded manifest (agent_label, job_description,
authority_scope_categories, supported_methods,
trust_score)
- Integrity hash of all package contents
- For .nomo: CA-signed certificate chain
2. SUBMIT (ACTIVATE transaction)
Registrant submits ACTIVATE request to governance endpoint:
- Package file (.agent or .nomo)
- Proposed agent label and optional org_domain or org_label
- Owner identity (maps to Agent Genesis owner field)
- Archetype declaration
- Declared trust_tier and verification_path with corresponding
tier-specific evidence (see {{AGTP-TRUST}} for the evidence
required by each verification path)
3. VALIDATE (governance platform)
Governance platform:
- Verifies package integrity hash
- For .nomo: validates certificate chain
- Applies verification-path-specific validation per
{{AGTP-TRUST}} (DNS challenge, transparency log submission,
or hybrid evidence verification, depending on declared path)
- Checks proposed label for uniqueness within the relevant
namespace (org_domain, org_label, or log-scoped)
4. ISSUE (Agent Genesis and canonical Agent-ID assigned)
Governance platform:
- Issues Agent Genesis with all fields populated,
including trust_tier and verification_path
- Derives canonical Agent-ID from certificate_hash
- For log-anchored Tier 1: submits Agent Genesis to
transparency log and embeds inclusion proof in the registry
record (see {{AGTP-LOG}})
- Creates registry record with Active lifecycle state
- Records genesis audit entry in immutable audit log
(genesis record includes full Agent Genesis and
verification evidence)
- Publishes agent to Namespace Document
(triggers Namespace Document re-signing)
The Agent Genesis is delivered to the registrant.
It is the permanent record of the agent's genesis.
Loss of the Agent Genesis does not invalidate the agent;
the certificate_hash remains the authoritative identity anchor.
5. ACTIVE
Agent enters Active lifecycle state.
Canonical Agent-ID is valid for AGTP protocol sessions.
All applicable alias URIs resolve to the Agent Identity
Document derived from the Agent Genesis.
6. LIFECYCLE EVENTS (post-activation)
SUSPEND: Agent temporarily inactive. Identity Document returns 503.
Agent Genesis and canonical ID remain valid.
Initiated by trust violation or human decision.
REINSTATE: Human-authorized return to Active state.
Agent Genesis unchanged. Reinstatement
recorded in audit trail.
REVOKE: Agent permanently deactivated. Identity Document returns 410.
Agent Genesis archived. Canonical ID retired
permanently and never reissued.
DEPRECATE: Controlled end-of-life. Identity Document returns 410 with
successor_agent field if applicable. Agent Genesis
retained per Section 8.5 retention policy.
6.5.8. Friendly Name Availability and Re-Registration
An agent label becomes available for re-registration 90 days after its associated agent enters Revoked or Deprecated lifecycle state. The canonical Agent-ID and Agent Genesis are permanently archived. The canonical Agent-ID MUST NOT be reissued under any circumstances, including re-registration of the same label by the same organization. This policy prevents ID reuse attacks in which a newly registered agent inherits the trust history of a revoked predecessor.¶
7. Method Definitions
7.1. Design Philosophy
AGTP methods are intent verbs, not resource operations. Each method expresses what an agent is trying to accomplish. Method names are uppercase ASCII strings. Methods that modify state are NOT idempotent by default unless explicitly marked. All methods accept a context parameter carrying agent session state. Requirement language follows [RFC2119].¶
7.1.1. The Twelve-Method Floor
AGTP defines twelve protocol-level methods that constitute the embedded floor of the protocol. Any conformant AGTP implementation MUST support all twelve. The twelve are organized as two groups of six.¶
Cognitive methods express requests for information or reasoning. A cognitive method invocation produces understanding without changing state external to the agent's own session:¶
-
QUERY: semantic data retrieval¶
-
DISCOVER: locate agents, resources, or services¶
-
DESCRIBE: retrieve operational capabilities of a known endpoint¶
-
SUMMARIZE: synthesize provided content¶
-
PLAN: produce an unexecuted sequence of actions toward a goal¶
-
PROPOSE: submit a dynamic endpoint or method proposal¶
Mechanics methods perform actions, exercise authority, or change state. A mechanics method invocation has external effect:¶
-
EXECUTE: invoke a specific action or carried protocol payload¶
-
DELEGATE: transfer execution to a sub-agent with bounded authority¶
-
ESCALATE: route a decision to a higher authority or human principal¶
-
CONFIRM: attest to a prior action, state, or item¶
-
SUSPEND: pause an active session workflow with a resumption nonce¶
-
NOTIFY: push information without expectation of synchronous response¶
The cognitive/mechanics distinction is informational, not normative; servers do not behave differently based on the category of a method. The classification provides a stable mental model for spec readers and implementers and constrains future additions to the floor.¶
EXECUTE is the generic carrier for application-layer payloads. When a higher-level agent framework such as MCP, A2A, or ACP is composed over AGTP (see Section 8), EXECUTE is the method that dispatches the carried payload to the application based on its Content-Type. EXECUTE absorbs the role that earlier drafts of this specification considered for an INVOKE method.¶
7.2. Core Methods
The twelve core methods are presented below in the cognitive group followed by the mechanics group. The four methods present in v06 core that are not included in the v07 floor (BOOK, SCHEDULE, LEARN, COLLABORATE) are demoted to Tier 2 standard extended methods; their specifications continue to apply where implementations choose to support them, and they are catalogued in the AGTP-API method catalog ([AGTP-API]).¶
7.2.1. QUERY
Purpose: Semantic data retrieval. The agent specifies what it needs to know, not where to find it. Distinguished from HTTP GET by expressing an information need rather than retrieving a known resource at a known location. Cognitive.¶
| Parameter | Required | Description |
|---|---|---|
| intent | MUST | Natural language or structured expression of the information need |
| scope | SHOULD | Data domains or sources to include or exclude |
| format | MAY | Desired response format: structured, natural, raw |
| confidence_threshold | MAY | Minimum confidence score for included results (0.0-1.0) |
| context | MAY | Session context for disambiguation |
Response: Result set with confidence scores per item. Server SHOULD indicate provenance of each result. Idempotent: Yes.¶
7.2.2. DISCOVER
Purpose: Locate agents, resources, or services matching specified criteria. Returns a list of candidate canonical Agent-IDs (or resource references) with minimal metadata for selection. Distinguished from QUERY: DISCOVER returns endpoints to talk to, QUERY returns data to consume. Cognitive.¶
| Parameter | Required | Description |
|---|---|---|
| criteria | MUST | Structured or natural-language description of what is being sought |
| filter | MAY | Constraints (capabilities, scopes, tier, geography, principal) |
| max_results | MAY | Maximum number of candidates to return |
| context | MAY | Session context for disambiguation |
Response: Ordered list of candidate entries, each containing
agent_id, name, description, principal, and a relevance score.
Server SHOULD apply governance-zone filtering: candidates outside
the requesting agent's permitted zones MUST NOT be returned.
Idempotent: Yes. Primary error codes: 422.¶
7.2.3. DESCRIBE
Purpose: Return the operational capabilities of a known agent endpoint.
The requesting agent specifies what capability dimensions it needs to
evaluate; the server returns a structured Capability Document. Used for
pre-task negotiation before committing to DELEGATE or EXECUTE. If
capability_domains is omitted, the server SHOULD return all supported
domains. Cognitive.¶
| Parameter | Required | Description |
|---|---|---|
| capability_domains | SHOULD | Comma-separated domains to return: methods, modalities, tools, version, budget, zones. If omitted, server SHOULD return all. |
| version_min | MAY | Minimum acceptable version for capability negotiation. |
| context | MAY | Session context for capability filtering. |
Response: Capability Document with the following structure:¶
{
"methods": ["QUERY", "DISCOVER", "DESCRIBE", "SUMMARIZE", "PLAN",
"PROPOSE", "EXECUTE", "DELEGATE", "ESCALATE", "CONFIRM",
"SUSPEND", "NOTIFY"],
"modalities": ["text", "image", "streaming"],
"tools": ["web_search", "code_execute"],
"version": "0.7",
"version_min_satisfied": true,
"trust_score": 0.94,
"budget_units_accepted": ["tokens", "compute-seconds"],
"zones_accepted": ["zone:internal", "zone:partner"]
}
¶
Idempotent: Yes. Primary error codes: 404, 422.¶
7.2.4. SUMMARIZE
Purpose: Request a concise synthesis of provided content or a referenced resource. The agent is requesting a cognitive operation on data, not retrieving data. Cognitive.¶
| Parameter | Required | Description |
|---|---|---|
| source | MUST | Content inline (up to implementation limit) or URI reference |
| length | SHOULD | Target summary length: brief, standard, detailed |
| focus | MAY | Aspect to emphasize in the summary |
| format | MAY | Output format: bullets, prose, structured |
| audience | MAY | Intended reader context, for calibrating complexity |
Response: Summary content with a source_hash and a confidence score. Idempotent: Yes.¶
7.2.5. PLAN
Purpose: Produce a sequence of actions or method invocations that would accomplish a stated goal, without executing them. The response is a structured plan that the requesting agent (or its principal) reviews before committing to execution. Distinguished from the deprecated v06 SCHEDULE method (which committed the plan): PLAN is purely cognitive and produces no external state change. Cognitive.¶
| Parameter | Required | Description |
|---|---|---|
| goal | MUST | Natural language or structured description of the desired outcome |
| constraints | SHOULD | Time, cost, scope, or policy constraints the plan must respect |
| available_methods | MAY | Methods the planning agent may assume the executor will support |
| context | MAY | Session context informing plan selection |
Response: Plan document containing ordered_steps (each an AGTP
method invocation specification), estimated_cost, confidence, and
assumptions. Idempotent: Yes. Primary error codes: 422, 503.¶
7.2.6. PROPOSE
Purpose: Submit a dynamic endpoint or method proposal to a server that has not yet declared support for the proposed method. PROPOSE initiates the dynamic endpoint negotiation flow described in Section 7.5. The proposing agent submits a method name, parameter signature, and intended outcome; the server responds with acceptance, counter-proposal, or rejection. Cognitive.¶
| Parameter | Required | Description |
|---|---|---|
| proposed_method | MUST | Uppercase ASCII method name being proposed |
| signature | MUST | Parameter set, response shape, and outcome semantics |
| intent | MUST | Natural language statement of what the method would accomplish |
| contract_version | SHOULD | Reference to the AGTP-API contract version (default AGTP-API/1.0) |
| context | MAY | Session context informing the proposal |
Response: Synthesis result. On accept, 263 Proposal Approved with
the AGTP-API endpoint definition, synthesis_id, and expires_at.
On refuse, 463 Proposal Rejected with a structured reason. On
async evaluation, 261 Negotiation In Progress with a proposal_id
the agent can poll via QUERY /proposals/{proposal_id} (see
[AGTP-API]). Idempotent: No. Primary error codes: 261, 262, 263,
400, 459, 460, 463.¶
7.2.7. EXECUTE
Purpose: Invoke a specific action with parameters, or dispatch a carried application-layer payload. EXECUTE is the protocol's generic mechanics carrier. When a higher-level agent framework is composed over AGTP, EXECUTE is the method that carries the framework's payload and dispatches it to the application based on the carried Content-Type. The Authority-Scope header is enforced against the action being performed; scope violations return 455. Mechanics.¶
| Parameter | Required | Description |
|---|---|---|
| action | MUST | Identifier of the action to invoke (application-defined when carrying a higher-level protocol) |
| parameters | SHOULD | Action-specific parameters (structured) |
| payload | MAY | Application-layer payload, when EXECUTE is carrying a higher-level protocol invocation |
| payload_type | MAY | Content-Type of the carried payload, for dispatch |
| idempotency_key | MAY | Client-provided key enabling safe retry |
| context | MAY | Session context |
Response: Execution result, with structure determined by the
action or carried payload. Servers MUST return the AGTP envelope
(status, task_id, attribution) wrapping the action's response.
Idempotent: depends on the action; clients SHOULD supply
idempotency_key for retry safety. Primary error codes: 403, 422,
455, 456, 503.¶
7.2.8. DELEGATE
Purpose: Transfer execution of a task or method to a sub-agent or downstream system. Initiates a new AGTP session on behalf of the delegating agent, carrying forward authority lineage. Mechanics.¶
| Parameter | Required | Description |
|---|---|---|
| target_agent_id | MUST | Identifier of the agent to delegate to |
| task | MUST | AGTP method call (or sequence) to execute |
| authority_scope | MUST | Scope granted to sub-agent MUST be a strict subset of delegating agent's scope |
| delegation_token | MUST | Signed token proving delegation authority |
| callback | SHOULD | AGTP endpoint for result delivery |
| deadline | MAY | Maximum time for task completion |
Security note: the authority_scope in a DELEGATE request MUST NOT
exceed the delegating agent's own Authority-Scope. Servers MUST enforce
this and MUST return 262 Authorization Required with body type
scope-required if violated. This is the protocol-level defense against
authority laundering. Idempotent: No.¶
7.2.9. ESCALATE
Purpose: Route a task, decision, or exception to a human principal or higher-authority agent when the current agent cannot or should not proceed. ESCALATE is the protocol-level expression of meaningful friction in AI systems as a first-class method. Mechanics.¶
| Parameter | Required | Description |
|---|---|---|
| task_id | MUST | The task or method invocation triggering escalation |
| reason | MUST | Structured reason: confidence_threshold, scope_limit, ethical_flag, ambiguous_instruction, resource_unavailable |
| context | MUST | Full context needed for the escalation recipient to act |
| priority | SHOULD | urgent, normal, or low |
| recipient | MAY | Specific human or agent to escalate to; if absent, routes to default handler |
| deadline | MAY | Time by which a response is needed |
Response: Escalation receipt with escalation_id and routing confirmation. The escalated task is paused until resolved via CONFIRM. Idempotent: Yes. An agent that escalates appropriately is functioning correctly. Governance frameworks built on AGTP can use escalation frequency and reason codes as observability signals for systemic issues.¶
7.2.10. CONFIRM
Purpose: Explicit acknowledgment of a prior action, state, or data item. Creates a signed attestation record. Mechanics.¶
| Parameter | Required | Description |
|---|---|---|
| target_id | MUST | ID of the action, booking, schedule, or item being confirmed |
| status | MUST | accepted, rejected, or deferred |
| reason | SHOULD (if rejected/deferred) | Explanation of the decision |
| attestation | MAY | Agent-signed confirmation payload for audit |
Response: Confirmation receipt with timestamp and attestation_id. Idempotent: Yes.¶
7.2.11. SUSPEND
Purpose: Pause a specific active session workflow in a recoverable state. Issues a resumption nonce the requesting agent uses to resume the session. Method-level SUSPEND is session-scoped and does not affect registry lifecycle state or Agent Genesis validity. The distinction between method-level SUSPEND and lifecycle SUSPEND (Section 6.7.6) is architectural: method-level SUSPEND is a workflow primitive; lifecycle SUSPEND is an administrative action on the agent's registry record. Mechanics.¶
| Parameter | Required | Description |
|---|---|---|
| session_id | MUST | The session to suspend. |
| reason | SHOULD | Structured reason: awaiting_input, resource_limit, scheduled_pause, external_dependency. |
| resume_by | MAY | ISO 8601 deadline for resumption. If exceeded without RESUME, session transitions to expired. |
| checkpoint | MAY | Agent-provided state snapshot for resumption context. Stored by server for duration of suspension. |
Response: Suspension receipt with the following structure:¶
{
"suspension_id": "susp-0042",
"session_id": "sess-a1b2c3d4",
"resumption_nonce": "[128-bit random value, base64url]",
"resume_by": "2026-04-15T09:00:00Z",
"status": "suspended"
}
¶
The resumption_nonce MUST be a cryptographically random 128-bit
value encoded as base64url. It is single-use: once presented to resume
a session, the nonce is invalidated and MUST NOT be accepted again.
Idempotent: No. Primary error codes: 404, 408.¶
Servers MUST generate nonces with at least 128 bits of entropy using a CSPRNG.¶
7.2.12. NOTIFY
Purpose: Asynchronous push of information from an agent to a recipient. Does not expect a response. Fire-and-forget. Delivery confirmation (if required) returned via a subsequent CONFIRM from the recipient. Mechanics.¶
| Parameter | Required | Description |
|---|---|---|
| recipient | MUST | Target Agent-ID, human endpoint, or broadcast group |
| content | MUST | Notification payload |
| urgency | SHOULD | critical, informational, or background |
| delivery_guarantee | MAY | at_most_once, at_least_once, or exactly_once |
| expiry | MAY | Timestamp after which the notification should not be delivered |
Response: Delivery receipt with notification_id. Idempotent: No.¶
7.3. Method Summary Table
| Method | Group | Intent | State-Modifying | Idempotent | Primary Error Codes |
|---|---|---|---|---|---|
| QUERY | Cognitive | Retrieve information | No | Yes | 404, 422 |
| DISCOVER | Cognitive | Locate agents or resources | No | Yes | 422 |
| DESCRIBE | Cognitive | Retrieve endpoint capabilities | No | Yes | 404, 422 |
| SUMMARIZE | Cognitive | Synthesize content | No | Yes | 400, 422 |
| PLAN | Cognitive | Produce an unexecuted plan | No | Yes | 422, 503 |
| PROPOSE | Cognitive | Submit a dynamic endpoint proposal | No | No | 261, 262, 263, 400, 459, 460, 463 |
| EXECUTE | Mechanics | Invoke an action or carried payload | Yes | Per action | 262, 403, 422, 455, 456, 503 |
| DELEGATE | Mechanics | Transfer task to sub-agent | Yes | No | 262, 403, 551 |
| ESCALATE | Mechanics | Defer to human or authority | Yes | Yes | 404 |
| CONFIRM | Mechanics | Attest to a prior action | Yes | Yes | 404, 400 |
| SUSPEND | Mechanics | Pause session workflow | Yes | No | 404, 408 |
| NOTIFY | Mechanics | Push information | No | No | 400, 404 |
7.4. Method Registry and Extensibility
AGTP defines a formal Method Registry maintained by IANA (see Section 8.2). Any party may submit a new method for registration. The registration procedure is Expert Review, and registration MUST be accompanied by a published specification, at minimum an IETF Internet-Draft or equivalent publicly available document. Registered methods MUST:¶
-
Have a unique uppercase ASCII name¶
-
Define required and optional parameters¶
-
Define expected response structure¶
-
Specify idempotency behavior¶
-
Specify applicable error codes¶
-
Include a security considerations section¶
-
Be accompanied by a published reference specification (Internet-Draft or RFC)¶
-
Appear in the AGTP-API approved method catalog [AGTP-API]. The verb list is the authoritative source for what method names are recognized by AGTP servers; registered methods are a subset of the method catalog with formal IETF registration. Submissions whose method name is not in the AGTP-API method catalog MUST be rejected by the Designated Expert.¶
Experimental methods MAY be used prior to registration using the X- prefix convention (e.g., X-NEGOTIATE). Experimental methods MUST NOT be used in production deployments without registration. Experimental methods are subject to verb-list validation; non-conformant experimental methods MUST be rejected with 459 Method Violation by AGTP-aware infrastructure components.¶
7.4.1. Verb-List Validation
AGTP servers MUST validate inbound method names against the AGTP-API approved method catalog. A method name not in the method catalog MUST result in a 459 Method Violation response. A method name in the method catalog whose path violates AGTP-API path grammar MUST result in a 460 Endpoint Violation response. A method name and path that are individually valid but not exposed by the server's policy MUST result in a 405 Method Not Allowed response, with the response body listing allowed methods for the path. The full contract validation model is specified in [AGTP-API].¶
Capability negotiation occurs during session establishment. The server returns a Supported-Methods header listing the methods it implements. Clients SHOULD check this list before invoking non-core methods.¶
The Negotiation-ID header is used to correlate turns within a dynamic
endpoint negotiation sequence (see Section 6.5). It MUST be a
UUID generated by the service upon receiving a PROPOSE request and
MUST be echoed in all subsequent turns of the same negotiation.
Maximum three turns before the agent MUST ESCALATE.¶
Negotiation-ID: 550e8400-e29b-41d4-a716-446655440000¶
QUOTE is defined as a Tier 2 Standard Extended Method in [AGTP-API].
QUOTE provides pre-flight cost estimation for a proposed method invocation:
the requesting agent submits a proposed method call; the server returns a
Cost-Estimate response without executing the method. Servers supporting
budget negotiation via the Budget-Limit header SHOULD implement QUOTE
to enable agents to validate cost before committing to execution. Servers
that implement QUOTE MUST list it in the Supported-Methods response
header at session establishment.¶
7.5. Dynamic Endpoint Negotiation
7.5.1. Overview
AGTP version 03 introduces a dynamic endpoint negotiation protocol enabling agents to discover data availability and instantiate endpoints on demand, without requiring pre-built API definitions. This protocol realizes the agentic API vision in which organizations expose data availability rather than pre-designed endpoints, and agents construct the interface they need at runtime.¶
The negotiation protocol operates at the transport layer. AGTP-API [AGTP-API] provides the contract structure for proposal and acceptance documents. The agent's identity and authority credentials (via the AGTP-CERT extension [AGTP-CERT] where deployed) govern authorization decisions.¶
7.5.2. Protocol Flow
Step 1: Pre-auth discovery
Agent issues unauthenticated GET to agtp://service.example.com
Service returns server manifest + data_manifest block
No credentials required at this step
Step 2: Agent evaluates data_manifest
Agent determines the service has relevant data
Agent assesses whether `negotiable: true` is declared
Agent constructs an AGTP-API-conformant endpoint proposal
Step 3: PROPOSE request
Agent sends PROPOSE with AGTP-API endpoint definition in body
Request MAY be unauthenticated if data sensitivity is low
Request MUST include AGTP-API/1.0 contract validation
Step 4a: Authorization required (262)
Service returns 262 with required authorization mechanism
Agent establishes credentials via specified mechanism
Agent resubmits PROPOSE with credentials
Negotiation-ID issued by service in 262 response
Step 4b: Negotiation in progress (261)
Service evaluates proposal asynchronously
Service returns 261 with Negotiation-ID
Agent polls or awaits outcome
Step 5a: Proposal approved (263)
Service returns 263 with complete AGTP-API endpoint definition,
synthesis_id, and expires_at
Instantiated endpoint is session-scoped by default
Agent MAY call the endpoint immediately
Step 5b: Proposal rejected (463)
Service returns 463 with structured reason (`out-of-scope`,
`policy-refused`, `composition-impossible`, `ambiguous`,
`synthesis-disabled`) and optional counter_proposal
Agent MAY modify proposal and retry (maximum 3 turns)
After 3 rejections agent MUST ESCALATE
¶
7.5.3. PROPOSE Method
PROPOSE is one of the twelve floor methods. The requesting agent submits an AGTP-API-conformant endpoint definition describing the interface it needs. The service evaluates whether it can fulfill the proposal against its endpoint primitives and authorization policy.¶
Parameters:¶
| Parameter | Required | Description |
|---|---|---|
| proposal | Yes | Complete AGTP-API endpoint definition (method + path + semantic block + input schema + output schema) |
| session_id | Yes | The active AGTP session identifier |
| persistent | Optional |
false (default) or true. Persistent syntheses survive across sessions, bounded by server policy. |
| requested_duration | Optional | ISO 8601 duration the agent requests for the synthesis (e.g., "24h", "7d"). Server policy bounds the granted duration. |
| scope_requested | Recommended | The Authority-Scope the agent requests for this endpoint |
| context | Optional | Session context informing the proposal |
Response on 263 Proposal Approved:¶
{
"negotiation_id": "550e8400-e29b-41d4-a716-446655440000",
"instantiated_endpoint": {
"method": "LOCATE",
"path": "/customer/{id}/location",
"semantic": {
"intent": "Returns the last known location for a customer",
"actor": "agent",
"outcome": "Location coordinates and address are returned",
"capability": "retrieval",
"confidence": 0.70,
"impact": "informational",
"is_idempotent": true
},
"input": { "required": ["id"] },
"output": { "coordinates": "object", "address": "string" },
"errors": ["customer_not_found", "location_not_available"],
"proposed": true,
"scope_required": "location:read",
"expires": "session"
}
}
¶
The proposed: true flag marks this as a dynamically instantiated
endpoint per the AGTP-API specification [AGTP-API].¶
7.5.4. Credential-Free Negotiation
For data classes declared with sensitivity: informational and
requires_authorization: false in the data manifest, services MAY
complete the full negotiation flow without requiring credentials.
The agent arrives, proposes, and receives an instantiated endpoint
without API keys.¶
For sensitive data classes, services MUST require credential establishment at Step 4a. The negotiation protocol is the mechanism by which credentials are established, not a prerequisite. This distinction is fundamental: the agent does not need credentials to begin a negotiation; it needs credentials to complete one for sensitive data.¶
AGTP-CERT [AGTP-CERT] provides the cryptographic identity binding that enables services to make fine-grained authorization decisions during negotiation based on the agent's verified identity, principal, and authority scope.¶
7.5.5. Session Scope and Persistence
Instantiated endpoints are session-scoped by default. They cease to exist when the AGTP session terminates. Services MAY offer persistent instantiation (the endpoint survives session termination and is added to the service's server manifest) subject to elevated authorization.¶
Persistent instantiation SHOULD be treated as a modification to the service's published server manifest. Services supporting persistent instantiation MUST increment their AGTP-API-Version header on the next discovery request following persistence.¶
7.6. Extended Method Vocabulary and Industry Profiles
7.6.1. Three-Tier Method Architecture
The AGTP method vocabulary is organized into three tiers reflecting different levels of universality, specificity, and domain relevance. All methods at all tiers MUST conform to the AGTP-API specification [AGTP-API]. The AGTP-API action-intent semantic class constraint applies to every method in the IANA registry and to every AGTP-API-validated custom method accepted via the AGTP-API method catalog.¶
- Tier 1. Core Methods (defined in Section 6.2):
-
The baseline vocabulary required for AGTP compliance. Every conformant AGTP implementation MUST support all Tier 1 methods. All Tier 1 methods are AGTP-API-conformant; they are defined instances of the action-intent semantic class standardized in [AGTP-API].¶
- Tier 2. Standard Extended Methods:
-
Registered in the IANA AGTP Method Registry and available for use in any AGTP implementation. Not required for baseline compliance but SHOULD be implemented where their semantics apply. Catalogued in the AGTP-API method catalog ([AGTP-API]). All Tier 2 methods satisfy AGTP-API contract requirements.¶
- Tier 3. Industry Profile Methods:
-
Domain-specific method sets defined and registered by industry communities as named AGTP profiles. Valid within deployments that declare support for the relevant profile. Not required in general-purpose implementations. All Tier 3 profile method submissions MUST include AGTP-API conformance verification as part of their specification.¶
- Tier 4. AGTP-API-Validated Custom Methods:
-
Organization-defined methods that are not registered in the IANA AGTP Method Registry but appear in the AGTP-API approved verb list and conform to AGTP-API path grammar. Valid within the deploying organization's AGTP services. The action-intent semantic class constraint applies identically. Agents discover and interpret these methods through natural language inference against AGTP-API semantic declarations, as validated empirically in [HOOD2026].¶
7.6.2. Method Category Taxonomy
All AGTP methods are organized into five categories:¶
- ACQUIRE:
-
Retrieve data, resources, or state without modifying it. Typically idempotent; no state modification.¶
- COMPUTE:
-
Process, transform, or analyze information and produce a derived result. Typically idempotent given the same input.¶
- TRANSACT:
-
Perform state-changing operations with external systems, resources, or records. Not idempotent by default; subject to reversibility classification.¶
- COMMUNICATE:
-
Send information, notifications, or signals to recipients. Fire-and- forget or confirm-receipt delivery models.¶
- ORCHESTRATE:
-
Coordinate, sequence, or manage multiple agents, tasks, or workflows. May spawn sub-agents or sessions; delegation chain semantics apply.¶
| Core Method | Group | Category |
|---|---|---|
| QUERY | Cognitive | Acquire |
| DISCOVER | Cognitive | Acquire |
| DESCRIBE | Cognitive | Acquire |
| SUMMARIZE | Cognitive | Compute |
| PLAN | Cognitive | Compute |
| PROPOSE | Cognitive | Orchestrate |
| EXECUTE | Mechanics | Transact |
| DELEGATE | Mechanics | Orchestrate |
| ESCALATE | Mechanics | Orchestrate |
| CONFIRM | Mechanics | Transact |
| SUSPEND | Mechanics | Orchestrate |
| NOTIFY | Mechanics | Communicate |
7.6.3. Standard Extended Methods (Tier 2)
The following methods constitute the initial Tier 2 registration set, defined in the AGTP-API method catalog ([AGTP-API]). Listed here by category with brief semantic definitions; full parameter specifications are in the catalog at the URL declared by AGTP-API.¶
ACQUIRE category: FETCH, SEARCH, SCAN, PULL, IMPORT, FIND.¶
COMPUTE category: EXTRACT, FILTER, VALIDATE, TRANSFORM, TRANSLATE, NORMALIZE, PREDICT, RANK, MAP.¶
TRANSACT category: REGISTER, SUBMIT, TRANSFER, PURCHASE, SIGN, MERGE, LINK, LOG, SYNC, PUBLISH.¶
COMMUNICATE category: REPLY, SEND, REPORT.¶
ORCHESTRATE category: MONITOR, ROUTE, RETRY, PAUSE, RESUME, RUN, CHECK.¶
Notable constraints: PURCHASE MUST carry explicit principal_id and scope enforcement; 455 Scope Violation applies if payments:purchase is not in the agent's Authority-Scope. RUN requires explicit procedure_id parameter; implementations MUST NOT accept free-form execution strings.¶
7.6.4. Short-Form and Industry-Inspired Methods
A set of short-form verb methods, e.g., SET, TAKE, OPEN, START, CALL, MAKE, TURN, BREAK, are provisionally catalogued as candidates for Tier 2 registration. These verbs are highly context-dependent and their semantics vary significantly across deployment domains.¶
Short-form methods will be registered individually only when a published companion specification provides unambiguous semantic definitions demonstrably distinct from existing registered methods. Provisional registrations using the X- prefix (e.g., X-SET, X-CALL) are encouraged during the experimentation period.¶
7.6.5. Industry Profile Method Sets
AGTP recognizes that specific industries require method vocabularies reflecting domain-specific operations that would be inappropriate in a general-purpose standard. Industry profile method sets are defined and registered as named AGTP profiles. A profile is a published companion specification that:¶
-
Declares a profile name (e.g., agtp-profile-healthcare, agtp-profile-financial, agtp-profile-legaltech)¶
-
Defines one or more industry-specific methods with full parameter specifications, error codes, and security considerations¶
-
Specifies which Tier 1 and Tier 2 methods are REQUIRED, RECOMMENDED, or NOT APPLICABLE within the profile¶
-
Addresses regulatory or compliance considerations specific to the domain (e.g., HIPAA for healthcare, PCI-DSS for financial services)¶
Illustrative examples of potential industry profile methods (not yet registered; listed for directional purposes only):¶
Healthcare: PRESCRIBE, AUTHORIZE, REFER, DISPENSE, TRIAGE, CONSENT, REDACT¶
Financial services: SETTLE, RECONCILE, HEDGE, CLEAR, UNDERWRITE, KYC, AML¶
Legal and compliance: ATTEST, NOTARIZE, DISCLOSE, REDLINE, EXECUTE, PRESERVE¶
Infrastructure: PROVISION, DEPROVISION, ROLLBACK, SNAPSHOT, FAILOVER¶
Industry communities are encouraged to develop and submit profile specifications through the IETF process. The IANA AGTP Method Registry will maintain a profile index alongside the core and standard method registries.¶
7.6.6. Registration Path for New Methods
- For Tier 2 Standard Methods:
-
Submit an Internet-Draft to the IETF providing full method specification per Section 6.4. The Designated Expert reviews for semantic uniqueness, clarity, AGTP-API contract conformance [AGTP-API], and security considerations. Submissions that fail AGTP-API contract validation MUST be returned to the submitter before review proceeds.¶
- For Industry Profile Methods (Tier 3):
-
Submit a profile specification to the IETF (or a recognized domain standards body with an established AGTP registry liaison) covering all methods in the profile and profile compliance requirements. The specification MUST include AGTP-API conformance statements for every method defined in the profile.¶
- For AGTP-API-Validated Custom Methods (Tier 4):
-
No IANA registration required. The implementing organization defines its method vocabulary in a server manifest as specified in [AGTP-API], served at the service's AGTP address. Methods are validated at the transport layer against the AGTP-API approved method catalog and AGTP-API path grammar. The method vocabulary is declared in the manifest's vocabulary block and discoverable by agents at runtime, with optional per-server vocabulary policy in the manifest's
policies.methodssub-block per [AGTP-API]. Organizations adopting Tier 4 methods are encouraged to publish their server manifests atagtp://[service-address]to enable cross-system agent discovery.¶ - For Experimental Methods:
-
Use the X- prefix without registration. Implementations MUST NOT deploy experimental methods in production without completing either the IANA registration process (Tier 2/3) or appearing in the AGTP-API approved method catalog (Tier 4). Experimental method names do not reserve the unprefixed name.¶
The AGTP Method Registry is published at: https://www.iana.org/assignments/agtp-methods/¶
The AGTP-API conformance test suite is maintained at: https://agtp.io/api/conformance¶
7.6.7. Real-time Service Adaptation
Services that update their server manifests at runtime MUST signal
changes via the AGTP-API-Version response header. This header MUST
be present on all AGTP responses from negotiable services.¶
AGTP-API-Version: 1.2.4¶
Agent runtimes MUST cache the AGTP-API-Version value from each service. When a response carries an AGTP-API-Version value different from the cached value, the agent runtime MUST re-fetch and re-validate the server manifest before issuing further method calls. This mechanism supports real-time service adaptation without requiring push notifications.¶
Adaptation flow: Agent calls EXECUTE for reserve-action Response includes AGTP-API-Version: 1.2.5 (was 1.2.4) Agent re-fetches agtp://service.example.com Service returns updated server manifest (new endpoint added) Agent updates service map Agent resumes operation with updated capability knowledge¶
Services SHOULD increment AGTP-API-Version when: - A new endpoint is added to the server manifest - An existing endpoint's semantic declaration changes - A new verb is added to the vocabulary block - A new data class is added to the data_manifest¶
Services MUST NOT decrement or reuse AGTP-API-Version values.¶
8. Composition with Higher-Level Frameworks
AGTP is not intended to replace MCP, A2A, ACP, ANP, or other agent application frameworks. AGTP is the substrate those frameworks need to operate at scale. The frameworks define what agents say to one another; AGTP defines how those messages move, who is allowed to send them, and how their effects are attributed.¶
This section establishes composition with higher-level frameworks as a first-class use case of AGTP, specifies the precedence rules between AGTP transport semantics and framework-level messaging semantics, and provides the canonical mapping table from common framework concepts onto AGTP primitives. Detailed composition profiles for each framework are specified in [AGTP-COMPOSITION].¶
8.1. Substrate Model
In a composed deployment, AGTP supplies four properties at the wire level that no framework currently supplies natively:¶
-
Identity at the protocol level. Every AGTP request carries a canonical Agent-ID in the
Agent-IDheader. The framework's payload need not carry identity; identity is established by the transport.¶ -
Authority bounded at the protocol level. The
Authority-Scopeheader declares what the requesting agent is permitted to do. Servers enforce scope before dispatching to the framework's payload handler. A scope violation returns 455 at the AGTP layer, before the framework sees the payload.¶ -
Attribution at the protocol level. Every method invocation produces an Attribution-Record. The framework's payload need not re-implement audit logging; the attribution is produced by the transport.¶
-
Delegation chain at the protocol level. Multi-hop agent workflows carry their lineage in the
Delegation-Chainheader, independent of framework-level concepts of "session" or "task."¶
A framework composed over AGTP can shed the parts of its design that exist only because HTTP did not provide these properties. The result is a framework that can focus on its actual contribution (tool semantics, conversational structure, multi-agent choreography) while the substrate handles the cross-cutting concerns.¶
8.2. EXECUTE as the Generic Carrier
When a framework is composed over AGTP, EXECUTE is the canonical method that carries the framework's payload. The carried payload's Content-Type identifies the framework; the AGTP server dispatches the payload to the framework's handler. See Section 7.2.7.¶
The EXECUTE invocation supplies:¶
-
The framework identifier (via
payload_type).¶ -
The framework-specific operation (via the
actionparameter).¶ -
The framework-specific payload (via the
payloadparameter).¶ -
An optional
idempotency_keyenabling safe retry.¶
The AGTP envelope (status, task_id, attribution) wraps the framework's response. A framework that wishes to expose a particular operation as a first-class AGTP method (rather than carrying it through EXECUTE) MAY define a Tier 4 AGTP-API-validated method per Section 7.1; this is reserved for frameworks whose operations are widely used enough to justify a dedicated verb.¶
8.3. Precedence Rule
AGTP headers (Agent-ID, Authority-Scope, Delegation-Chain, Session-ID, Task-ID) MUST take precedence over equivalent fields in a carried framework payload for routing, enforcement, audit, and identity purposes. Infrastructure components including SEPs and governance gateways MUST use AGTP header values for all protocol-level decisions.¶
Framework-level identity, session, or scope fields MAY be present in the body for application-layer use but MUST NOT override AGTP header values. If an inconsistency is detected between an AGTP header and a corresponding framework-level field, the AGTP header is authoritative; the inconsistency SHOULD be logged.¶
8.4. Canonical Mapping
| Framework | Concept | AGTP Mapping |
|---|---|---|
| MCP | Tool call | EXECUTE with payload_type: application/vnd.mcp.tools+json
|
| MCP | Resource fetch | QUERY with scope matching the resource's namespace |
| MCP | Sampling / inference | EXECUTE with payload_type identifying the sampling profile |
| MCP | Conversation context | Session-ID header carries the conversation across method invocations |
| A2A | Task | DELEGATE; A2A task.id maps to Task-ID header |
| A2A | Capability advertisement | DESCRIBE response |
| A2A | Agent Card | Agent Identity Document |
| A2A | Provenance chain | Delegation-Chain header |
| A2A | Artifact | NOTIFY body or EXECUTE response |
| ACP | Agent-to-agent message | NOTIFY (one-way) or EXECUTE (request-response) |
| ACP | Capability advertisement | DESCRIBE response |
| ANP | Identity exchange | Agent Identity Document; canonical Agent-ID |
| ANP | Discovery | DISCOVER method |
A wire example showing EXECUTE carrying an MCP tool invocation is provided in Appendix C. Additional examples for A2A and ACP are in the appendix-level composition section (Appendix F) and in [AGTP-COMPOSITION].¶
9. Merchant Identity and Agentic Commerce Binding
9.1. Overview
AGTP specifies agent-side identity through the Agent Birth
Certificate, canonical Agent-ID, Agent Identity Document, and Trust
Tier model defined in Section 5. PURCHASE invocations carrying
payments:purchase in the Authority-Scope header and a Budget-Limit
constraint are fully governed on the sending side. The receiving side
of a PURCHASE -- the merchant counterparty -- does not have an
equivalent protocol-level identity in the base specification.¶
Version 04 of AGTP introduces normative integration hooks for the AGTP Merchant Identity and Agentic Commerce Binding specification [AGTP-MERCHANT], which defines the merchant-side identity model. The integration is hook-based: this document registers the required headers, status code, and Authority-Scope domains; the detailed semantics, Merchant Manifest Document schema, Merchant Birth Certificate structure, and counterparty verification procedure are specified in the companion.¶
9.2. Merchant Identity Headers (Summary)
PURCHASE invocations in a fully conformant v04 deployment carry the following additional headers:¶
-
Merchant-ID: canonical identifier of the intended merchant counterparty.¶ -
Merchant-Manifest-Fingerprint: SHA-256 fingerprint of the Merchant Manifest Document the requesting agent verified during pre-flight counterparty verification.¶ -
Intent-Assertion: detached JWT carrying principal-authorized purchase intent, forwardable to payment networks as standalone evidence.¶ -
Cart-Digest: digest of a structured cart returned by a prior QUOTE invocation, binding this PURCHASE to that cart.¶
Full field definitions, wire examples, and security requirements are in [AGTP-MERCHANT].¶
9.3. 458 Counterparty Unverified (Summary)
Receiving servers MUST return 458 Counterparty Unverified on
PURCHASE invocations that fail merchant identity verification:
missing Merchant-ID or Merchant-Manifest-Fingerprint headers,
fingerprint mismatch, Merchant-ID mismatch, or a target merchant in
any lifecycle state other than Active. 455 is a governance signal,
parallel in role to 455 Scope Violation and 457 Zone Violation:
MUST be logged; MUST NOT be retried without re-running
counterparty verification.¶
9.4. Integration with PURCHASE, DISCOVER, and Attribution-Record
Three existing AGTP primitives interact with merchant identity:¶
-
PURCHASE: Counterparty verification runs before PURCHASE is sent on the wire. A verified PURCHASE produces an Attribution- Record naming both the agent and the merchant cryptographically.¶
-
DISCOVER: The DISCOVER method defined in [AGTP-DISCOVER] is extended by [AGTP-MERCHANT] to return Merchant Manifest Documents when the query carries
result_type: "merchant", and to return mixed agent/merchant result sets whenresult_type: "any". The existing DISCOVER signature model, ranking model, and governance-zone enforcement apply unchanged.¶ -
Attribution-Record: The Attribution-Record returned on PURCHASE includes
merchant_id,merchant_fingerprint, andintent_assertion_jtifields when merchant identity binding is in effect. This produces a dual-party cryptographic record consumable by downstream audit and dispute-resolution processes without requiring those processes to speak AGTP.¶
9.5. Relationship to Payment Networks
The merchant identity model defined in this document is payment-rail neutral. It does not define payment credential handling, tokenized card-on-file representations, authorization messaging to card networks, or settlement. Payment networks wishing to extend protection, fraud coverage, or dispute handling to agent-initiated transactions consume the Intent-Assertion JWT and the Attribution- Record as verifiable inputs to their own authorization and dispute flows; no AGTP-layer integration is required on the payment-network side. The specific mapping between AGTP merchant identity artifacts and payment-network message formats is expected to be defined bilaterally between governance platforms and individual networks and is out of scope for this document.¶
10. Security Considerations
This section satisfies the mandatory IETF Security Considerations requirement. All AGTP implementations MUST address the considerations described here.¶
10.1. Mandatory TLS
All AGTP connections MUST use TLS 1.3 or higher. Implementations MUST reject connections using TLS 1.2 or below. Certificate validation follows standard PKI practices per [RFC5280]. Servers MUST present a valid certificate.¶
10.2. Agent Identity Verification: Three Levels
AGTP defines three distinct levels at which agent identity and Authority-Scope can be verified. Each level serves a different deployment profile and operational tradeoff. Understanding the distinction is essential for implementers: the AGTP Agent Certificate extension ([AGTP-CERT]) is OPTIONAL, and base AGTP provides cryptographic verification at the application layer without it.¶
Level 1 - Self-asserted headers (raw request fields).
Every AGTP request MUST include the Agent-ID header, which
references the agent identity document carrying the principal
identifier and Authority-Scope. As a raw header value on an
individual request, Agent-ID is self-asserted: a client writes
the value into the request and the server records what was written.
Level 1 verification is limited to mandatory logging and anomaly
detection against the recorded stream. This is the minimum baseline
every AGTP implementation provides.¶
Level 2 - Application-layer cryptographic verification (signed Agent Identity Document). A canonical Agent-ID resolves to a signed Agent Identity Document (Section 5.5) that carries the Birth Certificate's Authority-Scope grant and is signed by the governance platform that issued it. A verifier (including a stranger with no prior relationship to the agent's organization) can cryptographically verify identity and scope at the application layer by performing the following steps:¶
-
Resolve the canonical Agent-ID to retrieve the signed Agent Manifest Document.¶
-
Verify the governance platform's signature on the manifest against the platform's published key.¶
-
Confirm that the canonical Agent-ID in the manifest matches the hash of the Agent Genesis.¶
-
Read the Authority-Scope grant from the verified manifest.¶
Level 2 verification is available in base AGTP without the Agent Certificate extension. It is the identity mechanism the protocol depends on. Self-asserted headers (Level 1) are bound to verified identity (Level 2) by the resolver's retrieval of the signed manifest for the declared canonical Agent-ID.¶
Level 3 - Transport-layer cryptographic verification (AGTP-CERT
extension). The AGTP Agent Certificate extension [AGTP-CERT]
binds the canonical Agent-ID and Authority-Scope to an X.509 v3
certificate presented during TLS 1.3 mutual authentication. The
principal identifier remains in the agent identity document
referenced by Agent-ID. Level 3 accelerates the Level 2 check to
the TLS handshake and enables Scope-Enforcement Points (SEPs) to
verify Authority-Scope at O(1) per-request cost without
application-layer access. Level 3 is an acceleration and
enforcement path for Level 2, not a replacement of it. Deployments
that require line-rate scope enforcement at infrastructure layers
(load balancers, governance gateways) SHOULD implement
[AGTP-CERT].¶
-
Note: The Agent Certificate extension and the Agent Genesis mechanism may be subject to pending intellectual property claims. See Section 7.7 and the IPR Notice preceding the Abstract for details. The licensor is prepared to grant a royalty-free license to implementers.¶
Every AGTP server MUST log the Agent-ID value from every
request and the principal identifier resolved from the
corresponding agent identity document, creating an attributable
audit trail at Level 1 even in deployments that do not implement
Level 2 retrieval or Level 3 transport binding.¶
10.4. Threat Model
10.4.1. Agent Spoofing
Threat: A malicious actor forges the Agent-ID header to
impersonate a trusted agent. Mitigation: Level 2 application-layer
verification binds a declared Agent-ID to the signed Agent Identity
Document retrieved via canonical ID resolution. A forged Agent-ID
either fails to resolve or resolves to a manifest whose signature
cannot be verified against the claimed governance platform's published
key. Level 3 raises the mitigation to the TLS handshake via
[AGTP-CERT]. Implementations SHOULD retrieve and verify the
manifest for any Agent-ID carrying scope beyond read-only query
operations. Mandatory Level 1 logging provides an anomaly-detection
baseline for deployments that do not perform active verification on
every request.¶
10.4.3. Delegation Chain Poisoning
Threat: A malicious agent inserts itself into a delegation chain. Mitigation: each hop of a DELEGATE chain MUST be logged with the delegating agent's Agent-ID, the sub-agent's Agent-ID, and the declared Authority-Scope. Servers processing a delegated request MUST be able to reconstruct the delegation sequence from log data, sub-agent identity documents, and Authority-Scope subset verification. 551 Authority Chain Broken is returned when any link in the sequence is unverifiable. Full mitigation requires [AGTP-CERT] for signed delegation tokens.¶
10.4.4. Denial of Service via High-Frequency Agent Traffic
Threat: Agents that are compromised, misconfigured, or adversarial generate extremely high request volumes. Mitigation: 429 Rate Limited status code. Rate limiting SHOULD be applied per Agent-ID and per resolved principal identifier (obtained from the agent identity document referenced by Agent-ID). When [AGTP-CERT] is deployed, per-Agent-ID quotas can be cryptographically tied to verified identity, preventing quota evasion through Agent-ID spoofing.¶
10.4.5. Session Hijacking
Threat: An attacker intercepts or forges a Session-ID. Mitigation: mandatory TLS protects sessions in transit. Session-IDs MUST be cryptographically random with minimum 128 bits of entropy. Servers MUST validate that Session-ID, Agent-ID, and TLS client identity are consistent.¶
10.4.6. Escalation Suppression
Threat: A compromised agent or intermediary suppresses ESCALATE requests, preventing human oversight. Mitigation: compliant implementations MUST route ESCALATE requests directly to the declared escalation handler without modification. Intermediaries MUST NOT drop, delay, or modify ESCALATE requests. Escalation handlers SHOULD implement independent receipt confirmation.¶
10.4.7. Agent Genesis Spoofing
Threat: A malicious actor fabricates an Agent Genesis to claim a
legitimate agent's identity or construct a false identity with elevated
trust. Mitigation: Agent Genesis documents are issued only by governance
platforms that have completed one of the three Tier 1 verification
paths (Section 5.2). For dns-anchored registrations, the governance
platform MUST verify DNS ownership of the claimed org_domain
before issuance. For log-anchored registrations, the governance
platform MUST submit the Agent Genesis to a transparency log
per [RFC9162] / [RFC9943] and record the inclusion proof in the
registry; tampering with a log-anchored Agent Genesis is
detectable by any party with log access. For hybrid registrations,
both DNS and blockchain address ownership are verified. In the base
spec, mandatory logging provides auditability. Full mitigation
requires [AGTP-CERT] for cryptographically bound Agent Genesis
verification at the transport layer. Governance platforms MUST
treat any ACTIVATE request that presents a certificate hash matching
an existing registry record as a collision attack and MUST reject
it.¶
10.4.8. Domain Transfer Identity Hijacking
Threat: An attacker acquires an expired domain to inherit the agent
registry and trust history of prior registrants. Mitigation applies
to dns-anchored and hybrid Tier 1 agents: agents under an expired
domain are automatically Suspended within 24 hours of domain expiry
detection. A new owner of the domain MUST NOT inherit prior agent
registrations. See Section 9.6 for the full domain expiry policy.
log-anchored Tier 1 agents are unaffected by this threat because
their verification evidence is the transparency log inclusion proof
rather than DNS ownership.¶
10.4.9. Attribution Forgery
Threat: A malicious agent submits a fabricated or replayed Attribution-Record to claim credit for an action it did not perform, or to conceal the true execution context of an action it did perform.¶
Mitigation: Attribution-Records MUST be signed with the agent's
governance key. The signature MUST cover the full record including the
Task-ID, Agent-ID, method, timestamp, and result hash. When [AGTP-CERT]
is deployed, the signature is verified at the transport layer against the
agent's X.509 certificate. For high-stakes domains, RATS attestation
evidence in the Attribution-Record per [RFC9334] provides hardware-rooted
proof of execution context that cannot be forged without compromising the
attesting environment itself. Attribution-Record signatures MUST be
verified before the record is admitted to an audit trail. Unverified records
MUST be logged with a signature_unverified flag and MUST NOT be
treated as authoritative for compliance purposes.¶
10.5. Privacy Considerations
Agent identity carried on requests, and the agent identity
document referenced by Agent-ID, carry information about agent
behavior that may be sensitive:¶
-
Agent-ID together with the resolved principal identifier may reveal organizational structure¶
-
Session-ID and Task-ID reveal workflow patterns¶
-
Delegation-Chain (reserved for future revisions) would reveal multi-agent architecture¶
AGTP logs containing these fields MUST be treated as sensitive operational data. Operators MUST implement appropriate access controls, retention limits, and data minimization practices consistent with applicable privacy regulations.¶
Where privacy-preserving attribution is required, implementations MAY use pseudonymous Agent-IDs with a separate trusted resolution service. The architecture for pseudonymous agent identity resolution is reserved for a future companion document.¶
10.6. Denial-of-Service Considerations
AGTP's agent identity provides a mechanism for more precise denial-of-service mitigation than is possible with HTTP. Rate limiting SHOULD be applied per Agent-ID and per resolved principal identifier (obtained from the agent identity document) in addition to per-IP-address controls.¶
When [AGTP-CERT] is deployed, per-Agent-ID rate limiting can be cryptographically tied to verified agent identity, preventing quota evasion through Agent-ID rotation. Implementations planning high-volume governed agent deployments SHOULD plan for [AGTP-CERT] as part of their denial-of-service mitigation strategy.¶
Additional recommended mitigations: traffic-shaping by request
class once the Priority header is normatively specified (reserved
for v01+ per Section 5.5); and circuit breaker patterns for
ESCALATE request floods.¶
10.7. Intellectual Property Considerations
The core AGTP specification, including all base methods, header fields, status codes, connection model, and IANA registrations defined in this document, is intended for open implementation without royalty obligation.¶
Certain elements referenced in this document may be subject to pending patent applications by the author, specifically:¶
-
The Agent Certificate extension [AGTP-CERT], which provides cryptographic binding of agent identity and authority scope to AGTP header fields.¶
-
The ACTIVATE method, which provides AGTP-native transmission and activation of governed agent packages.¶
-
The Agent Genesis mechanism (Section 5.7), which provides the genesis identity record and canonical Agent-ID derivation process for AGTP-registered agents.¶
-
The
.agentfile format specification, an open packaging format for AI agents.¶ -
The
.nomofile format specification, a governed packaging format for AI agents with cryptographic governance binding.¶
Implementers of the core AGTP specification are not affected by any intellectual property claims on these extensions and associated formats.¶
The licensor is prepared to grant a royalty-free license to implementers for any patent claims that cover contributions in this document and its referenced extensions, consistent with the IETF's IPR framework under [RFC8179].¶
IPR disclosures have been filed with the IETF Secretariat and are available at: https://datatracker.ietf.org/ipr/¶
11. IANA Considerations
This document records the following IANA registrations.¶
11.1. Port Assignment
The following service names and port number are registered in the IANA Service Name and Transport Protocol Port Number Registry:¶
| Service Name | Port | Transport | Description |
|---|---|---|---|
| agtp | 4480 | TCP | Agent Transfer Protocol over TCP/TLS |
| agtp-quic | 4480 | UDP | Agent Transfer Protocol over QUIC |
The unified port assignment (4480 for both TCP and UDP under a
single conceptual agtp service) follows the precedent set by HTTPS
(443/TCP and 443/UDP for HTTP/3). The transport is distinguished at
the protocol level rather than at the port number.¶
11.2. AGTP Method Registry
Establishment of an IANA registry: Agent Transfer Protocol Methods.¶
- Registry name:
-
Agent Transfer Protocol Methods¶
- Registration procedure:
-
Expert Review per [RFC8126], with the additional requirement that each registration be accompanied by a published specification, at minimum a publicly available Internet-Draft or equivalent document. The Designated Expert SHOULD verify that the proposed method name is unique, the reference specification is publicly accessible, the method definition includes the required fields (parameters, response structure, idempotency, error codes, security considerations), and the method conforms to the AGTP-API specification [AGTP-API].¶
- Reference:
-
This document¶
Initial registrations (the twelve-method protocol floor):¶
| Method | Group | Status | Reference |
|---|---|---|---|
| QUERY | Cognitive | Permanent | Section 7.2 |
| DISCOVER | Cognitive | Permanent | Section 7.2 |
| DESCRIBE | Cognitive | Permanent | Section 7.2 |
| SUMMARIZE | Cognitive | Permanent | Section 7.2 |
| PLAN | Cognitive | Permanent | Section 7.2 |
| PROPOSE | Cognitive | Permanent | Section 7.2 |
| EXECUTE | Mechanics | Permanent | Section 7.2 |
| DELEGATE | Mechanics | Permanent | Section 7.2 |
| ESCALATE | Mechanics | Permanent | Section 7.2 |
| CONFIRM | Mechanics | Permanent | Section 7.2 |
| SUSPEND | Mechanics | Permanent | Section 7.2 |
| NOTIFY | Mechanics | Permanent | Section 7.2 |
The methods BOOK, SCHEDULE, LEARN, and COLLABORATE, present in the v06 core set, are demoted in v07 to Tier 2 standard extended methods and are registered through the AGTP-API method catalog ([AGTP-API]) rather than through this document.¶
11.3. AGTP Status Code Registry
Establishment of an IANA registry: Agent Transfer Protocol Status Codes.¶
- Registry name:
-
Agent Transfer Protocol Status Codes¶
- Registration procedure:
-
Expert Review + published specification required.¶
AGTP-specific status code numbers are deliberately chosen from ranges unassigned in the IANA HTTP Status Code Registry to avoid semantic collision with HTTP status codes that may appear in payloads carried by AGTP method invocations.¶
Two status codes used by AGTP retain their HTTP code numbers (408 and 410) but carry AGTP-specific semantics. They are registered here with text describing the AGTP semantic.¶
The following AGTP status codes are registered with full definitions:¶
| Code | Name | Definition | Reference |
|---|---|---|---|
| 261 | Negotiation In Progress | The service has received a PROPOSE request and is evaluating it asynchronously. The response body MUST include a proposal_id and an estimated evaluation duration. The agent retrieves the terminal status by invoking QUERY /proposals/{proposal_id} until a 263 (Proposal Approved) or 463 (Proposal Rejected) response is returned. Server policy controls whether async evaluation is offered. |
[AGTP-API] |
| 262 | Authorization Required | The request requires credential establishment, additional authorization scope, or consent that is not yet present. The response body MUST specify which authorization condition applies: scope-required (the endpoint requires Authority-Scope the agent has not declared), wildcards-required (the request is an ad-hoc method invocation and the wildcards consent on either the agent identity document or the server policy is absent), credentials-missing (the server requires credentials such as AGTP-CERT or OAuth-scoped token before evaluating the request), or anonymous-discovery-disabled (the server requires authenticated identity for manifest retrieval and the request is unauthenticated). Returned for PROPOSE-time authorization, endpoint-dispatch scope checks, ad-hoc invocation wildcards refusal, and discovery requests blocked by policy. |
[AGTP-API] |
| 263 | Proposal Approved | The service has accepted the PROPOSE request and instantiated the proposed endpoint. The response body MUST contain a complete AGTP-API endpoint definition for the instantiated endpoint, a synthesis_id identifying the synthesized endpoint for follow-on invocation, and expires_at indicating when the synthesis expires. MAY include a persistent boolean and granted_duration indicating the actual duration the server granted (which may be less than the agent requested, bounded by server policy). |
[AGTP-API] |
| 405 | Method Not Allowed | The method is recognized and the path is valid, but the server's policy or registry does not expose this combination. The response body MUST list allowed methods for the path and any redirects from the manifest's policies.methods sub-block. The agent MAY PROPOSE the combination if it is not exposed by policy. |
[AGTP-API] |
| 408 | Timeout | The method's declared TTL expired before execution completed. AGTP-specific semantics distinct from HTTP's request-timeout: applies to AGTP method TTL rather than transport request timeout. MUST be logged. | Section 5.6 |
| 410 | Gone | The Agent-ID is permanently retired through revocation or deprecation of the underlying Agent Genesis. AGTP-specific semantics distinct from HTTP's resource-removed: applies to permanent identity retirement. The canonical Agent-ID MUST NOT be retried. | Section 5.6 |
| 455 | Scope Violation | The requested action is outside a declared scope dimension other than Authority-Scope, rate-limit, budget, or zone (which have dedicated codes 262, 429, 456, 457 respectively). Typical applications include token-based scope violations and query-based scope violations where the operator defines a scope dimension outside the standard set. The server MUST log this event. The agent MUST NOT retry the same request without modifying its scope declaration. Governance signal, not a protocol error. | Section 5.6 |
| 456 | Budget Exceeded | The requested method execution would exceed the resource limits declared in the Budget-Limit request header. The agent MUST NOT retry without modifying the Budget-Limit or reducing request scope. Governance signal; MUST be logged. | Section 5.6 |
| 457 | Zone Violation | The request would route outside the network boundary declared in the AGTP-Zone-ID header. SEP-enforced. The agent MUST NOT retry without modifying the AGTP-Zone-ID or obtaining explicit cross-zone authorization. MUST be logged. | Section 5.6 |
| 458 | Counterparty Unverified | The merchant counterparty in a PURCHASE invocation failed identity verification. Returned when the Merchant-ID or Merchant-Manifest-Fingerprint request headers are absent, when the fingerprint does not match the receiving server's current Merchant Manifest Document, when the Merchant-ID does not match the server's canonical ID, or when the merchant is in a non-Active lifecycle state. Governance signal; MUST be logged. Full definition in [AGTP-MERCHANT]. |
[AGTP-MERCHANT] |
| 459 | Method Violation | The method name is not in the AGTP-API approved method catalog. The method itself is the problem. The response body MUST identify the unrecognized method and SHOULD reference the AGTP-API method catalog version in effect. The agent MUST select a different method before retrying, or MAY PROPOSE the method if no suitable replacement exists. | [AGTP-API] |
| 460 | Endpoint Violation | The endpoint path violates AGTP-API path grammar. A path segment matches an approved method name (case-insensitive), indicating method-name leakage into the path. The response body MUST identify the offending path segment. The agent MUST restructure the path before retrying. | [AGTP-API] |
| 461 | Reserved | Reserved for AGTP expansion. | Section 5.6 |
| 462 | Reserved | Reserved for AGTP expansion. | Section 5.6 |
| 463 | Proposal Rejected | The service cannot or will not instantiate the proposed endpoint. Returned in response to PROPOSE. The response body MUST include a structured reason field with one of the values out-of-scope, policy-refused, composition-impossible, ambiguous, or synthesis-disabled, an explanation string, and MAY include an optional counter_proposal suggesting a related endpoint the server would accept. |
[AGTP-API] |
| 464 | Reserved | Reserved for AGTP expansion. | Section 5.6 |
| 465 | Reserved | Reserved for AGTP expansion. | Section 5.6 |
| 550 | Delegation Failure | A sub-agent to which a task was delegated via the DELEGATE method failed to complete the task within the declared deadline or returned an error. The response body SHOULD contain the sub-agent's error details. | Section 5.6 |
| 551 | Authority Chain Broken | One or more entries in the delegation sequence cannot be verified as part of a valid and continuous delegation chain. The specific unverifiable entry SHOULD be identified in the response body. The server MUST log this event. | Section 5.6 |
| 552 | Reserved | Reserved for AGTP expansion. | Section 5.6 |
| 553 | Reserved | Reserved for AGTP expansion. | Section 5.6 |
| 554 | Reserved | Reserved for AGTP expansion. | Section 5.6 |
| 555 | Reserved | Reserved for AGTP expansion. | Section 5.6 |
11.4. Media Type Registry
This section is the master registry of all AGTP-family media types. Types defined by companion specifications are listed here for cross-document discoverability; their normative definitions remain in the defining document.¶
| Media Type | Use | Defining Reference | IANA Status |
|---|---|---|---|
application/vnd.agtp+json
|
AGTP method request/response bodies (JSON) | Section 5.7 | Planned (this document) |
application/vnd.agtp+yaml
|
AGTP method request/response bodies (YAML) | Section 5.7 | Planned (this document) |
application/vnd.agtp.identity+json
|
Agent Identity Document (JSON) | Section 6.3 | Vendor-tree registration submitted |
application/vnd.agtp.identity+yaml
|
Agent Identity Document (YAML) | Section 6.3 | Vendor-tree registration submitted |
application/vnd.agtp.manifest+json
|
AGTP server manifest | [AGTP-API] | Planned (AGTP-API) |
application/vnd.agtp.endpoint+json
|
AGTP-API endpoint definition | [AGTP-API] | Planned (AGTP-API) |
"Vendor-tree registration submitted" indicates that the
registration application has been filed with IANA against the
application/vnd.* vendor tree and is awaiting processing.
"Planned (this document)" and "Planned (AGTP-API)" indicate types
for which registration applications will be filed concurrent with
publication of the defining document. Standards-tree promotion of
the full set is anticipated at RFC publication of the AGTP family.¶
11.5. Header Field Registry
AGTP header fields are distinct from HTTP header fields and are registered in a new IANA registry: Agent Transfer Protocol Header Fields.¶
- Registry name:
-
Agent Transfer Protocol Header Fields¶
- Registration procedure:
-
Expert Review + published specification required.¶
AGTP does not reuse the HTTP Field Name Registry, as AGTP header fields have different semantics, applicability, and versioning constraints from HTTP fields. HTTP header fields are not automatically valid in AGTP, and AGTP header fields are not valid HTTP fields.¶
Initial registrations (all Permanent): Agent-ID, Authority-Scope, Session-ID, Task-ID, Delegation-Chain, Server-ID, Attribution-Record, Continuation-Token, Supported-Methods, Cost-Estimate, Attestation-Evidence, Merchant-ID, Merchant-Manifest-Fingerprint, Intent-Assertion, Cart-Digest, AGTP-API-Version, AGTP-Catalog-Warning, AGTP-Endpoint-Warning.¶
The four merchant-related headers are defined in [AGTP-MERCHANT] and
registered concurrently with this document. The three API-related
headers (AGTP-API-Version, AGTP-Catalog-Warning, and
AGTP-Endpoint-Warning) are defined in [AGTP-API] and registered
concurrently.¶
Headers reserved for future revisions (Priority, TTL, Budget-Limit, AGTP-Zone-ID, Content-Schema, Telemetry-Export) are not registered in this revision. They will be registered alongside the future revision that specifies their normative semantics.¶
11.6. URI Scheme Registration
The agtp:// URI scheme is registered per [RFC7595]. Full
registration template is documented in Section 5.1.8 of this
document.¶
11.7. AGTP Budget Unit Registry
Establishment of a new IANA sub-registry: Agent Transfer Protocol Budget Units.¶
- Registry name:
-
Agent Transfer Protocol Budget Units¶
- Registration procedure:
-
Expert Review per [RFC8126]. New unit registrations MUST define: unit name (lowercase ASCII, no spaces or special characters), semantic description, value format (integer or decimal), whether fractional values are permitted, and a reference specification. Units representing financial denominations MUST specify the currency and MUST define precision (decimal places). The Designated Expert SHOULD verify that the proposed unit does not duplicate an existing registration and that the value format is unambiguous.¶
- Reference:
-
This document¶
Initial registrations:¶
| Unit | Description | Value Format | Fractional |
|---|---|---|---|
| tokens | Language model token consumption | Integer | No |
| compute-seconds | CPU/GPU compute time in seconds | Decimal | Yes |
| USD | US Dollar financial limit | Decimal, 2 places | Yes |
| EUR | Euro financial limit | Decimal, 2 places | Yes |
| GBP | Pound Sterling financial limit | Decimal, 2 places | Yes |
| calls | Downstream API call count | Integer | No |
11.8. Agent Registry Retention Policy
The AGTP registry MUST retain records for all registered agents regardless of lifecycle state. The following minimum retention periods apply:¶
| Lifecycle State | Minimum Retention Period |
|---|---|
| Active | Duration of Active state + 7 years |
| Suspended | Duration of Suspended state + 7 years |
| Revoked | 10 years from revocation date |
| Deprecated | 7 years from deprecation date |
The 7-year minimum reflects common enterprise compliance requirements (SOX, GDPR audit trails, HIPAA). Governance platform operators in regulated industries SHOULD extend these minimums to match applicable regulatory requirements.¶
The retained record for a Revoked or Deprecated agent MUST include:¶
-
Canonical Agent-ID (permanently retired, not reissued)¶
-
Agent label and org domain at time of registration¶
-
Trust tier at time of registration¶
-
Activation date and activating principal¶
-
Revocation or deprecation date, initiating principal, and reason code¶
-
Genesis audit record hash (pointer to immutable audit log)¶
-
Full Agent Genesis (archived, not publicly accessible)¶
-
All lifecycle state transitions with timestamps¶
The retained record MUST NOT contain package executable contents, active session data, or Authority-Scope grant tokens.¶
11.8.1. Domain Name Expiry Interaction
If an organization's org_domain expires or transfers to a new owner:¶
-
All Active agents registered under the expired domain MUST be automatically Suspended within 24 hours of domain expiry detection.¶
-
The governance platform MUST notify the registered principal contact before suspension takes effect, with a minimum notice period of 30 days if domain expiry was predictable.¶
-
Suspended agents under an expired domain transition to Deprecated state after 90 days if the domain has not been renewed.¶
-
A new owner of the domain MUST NOT inherit prior agent registrations. New ACTIVATE transactions are required.¶
This policy prevents domain-transfer-based identity hijacking in which an attacker acquires an expired domain to claim the trust history of agents that operated under it.¶
12. References
12.1. Normative References
- [AGTP-API]
- Hood, C., "AGTP-API: Verbs, Paths, Endpoints, and Synthesis", Work in Progress, Internet-Draft, draft-hood-agtp-api-00, , <https://datatracker.ietf.org/doc/html/draft-hood-agtp-api-00>.
- [RFC2119]
- Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, , <https://www.rfc-editor.org/rfc/rfc2119>.
- [RFC3986]
- Berners-Lee, T., Fielding, R., and L. Masinter, "Uniform Resource Identifier (URI): Generic Syntax", STD 66, RFC 3986, DOI 10.17487/RFC3986, , <https://www.rfc-editor.org/rfc/rfc3986>.
- [RFC5280]
- Cooper, D., Santesson, S., Farrell, S., Boeyen, S., Housley, R., and W. Polk, "Internet X.509 Public Key Infrastructure Certificate and Certificate Revocation List (CRL) Profile", RFC 5280, DOI 10.17487/RFC5280, , <https://www.rfc-editor.org/rfc/rfc5280>.
- [RFC6335]
- Cotton, M., Eggert, L., Touch, J., Westerlund, M., and S. Cheshire, "Internet Assigned Numbers Authority (IANA) Procedures for the Management of the Service Name and Transport Protocol Port Number Registry", BCP 165, RFC 6335, DOI 10.17487/RFC6335, , <https://www.rfc-editor.org/rfc/rfc6335>.
- [RFC8126]
- Cotton, M., Leiba, B., and T. Narten, "Guidelines for Writing an IANA Considerations Section in RFCs", BCP 26, RFC 8126, DOI 10.17487/RFC8126, , <https://www.rfc-editor.org/rfc/rfc8126>.
- [RFC8174]
- Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, , <https://www.rfc-editor.org/rfc/rfc8174>.
- [RFC8179]
- Bradner, S. and J. Contreras, "Intellectual Property Rights in IETF Technology", BCP 79, RFC 8179, DOI 10.17487/RFC8179, , <https://www.rfc-editor.org/rfc/rfc8179>.
- [RFC8555]
- Barnes, R., Hoffman-Andrews, J., McCarney, D., and J. Kasten, "Automatic Certificate Management Environment (ACME)", RFC 8555, DOI 10.17487/RFC8555, , <https://www.rfc-editor.org/rfc/rfc8555>.
- [RFC8615]
- Nottingham, M., "Well-Known Uniform Resource Identifiers (URIs)", RFC 8615, DOI 10.17487/RFC8615, , <https://www.rfc-editor.org/rfc/rfc8615>.
- [RFC9000]
- Iyengar, J., Ed. and M. Thomson, Ed., "QUIC: A UDP-Based Multiplexed and Secure Transport", RFC 9000, DOI 10.17487/RFC9000, , <https://www.rfc-editor.org/rfc/rfc9000>.
- [RFC9001]
- Thomson, M., Ed. and S. Turner, Ed., "Using TLS to Secure QUIC", RFC 9001, DOI 10.17487/RFC9001, , <https://www.rfc-editor.org/rfc/rfc9001>.
- [RFC9110]
- Fielding, R., Ed., Nottingham, M., Ed., and J. Reschke, Ed., "HTTP Semantics", STD 97, RFC 9110, DOI 10.17487/RFC9110, , <https://www.rfc-editor.org/rfc/rfc9110>.
- [RFC9334]
- Birkholz, H., Thaler, D., Richardson, M., Smith, N., and W. Pan, "Remote ATtestation procedureS (RATS) Architecture", RFC 9334, DOI 10.17487/RFC9334, , <https://www.rfc-editor.org/rfc/rfc9334>.
12.2. Informative References
- [A2A]
- Linux Foundation, "Agent-to-Agent Protocol Specification", , <https://a2aprotocol.ai>.
- [ACP]
- IBM Research, "Agent Communication Protocol", .
- [AGTP-CERT]
- Hood, C., "AGTP Agent Certificate Extension", Work in Progress, Internet-Draft, draft-hood-agtp-agent-cert-00, , <https://datatracker.ietf.org/doc/html/draft-hood-agtp-agent-cert-00>.
- [AGTP-COMPOSITION]
- Hood, C., "AGTP Composition with Agent Group Messaging Protocols", Work in Progress, Internet-Draft, draft-hood-agtp-composition-00, , <https://datatracker.ietf.org/doc/html/draft-hood-agtp-composition-00>.
- [AGTP-DISCOVER]
- Hood, C., "AGTP Agent Discovery and Name Service", Work in Progress, Internet-Draft, draft-hood-agtp-discovery-00, , <https://datatracker.ietf.org/doc/html/draft-hood-agtp-discovery-00>.
- [AGTP-LOG]
- Hood, C., "AGTP Transparency Log Protocol", Work in Progress, Internet-Draft, draft-hood-agtp-log-00, , <https://datatracker.ietf.org/doc/html/draft-hood-agtp-log-00>.
- [AGTP-MERCHANT]
- Hood, C., "AGTP Merchant Identity and Agentic Commerce Binding", Work in Progress, Internet-Draft, draft-hood-agtp-merchant-identity-00, , <https://datatracker.ietf.org/doc/html/draft-hood-agtp-merchant-identity-00>.
- [AGTP-TRUST]
- Hood, C., "AGTP Trust and Verification Specification", Work in Progress, Internet-Draft, draft-hood-agtp-trust-00, , <https://datatracker.ietf.org/doc/html/draft-hood-agtp-trust-00>.
- [AGTP-WEB3]
- Hood, C., "AGTP Web3 Bridge Specification", Work in Progress, Internet-Draft, draft-hood-agtp-web3-bridge-00, , <https://datatracker.ietf.org/doc/html/draft-hood-agtp-web3-bridge-00>.
- [ANP]
- "Agent Network Protocol", .
- [HOOD2026]
- Hood, C., "Semantic Method Naming and LLM Agent Accuracy: A Controlled Benchmark of REST/CRUD versus Agentive API Interface Design", Working Paper Available by request. March 2026., .
- [MCP]
- Anthropic, "Model Context Protocol", , <https://modelcontextprotocol.io>.
- [RFC6749]
- Hardt, D., Ed., "The OAuth 2.0 Authorization Framework", RFC 6749, DOI 10.17487/RFC6749, , <https://www.rfc-editor.org/rfc/rfc6749>.
- [RFC7519]
- Jones, M., Bradley, J., and N. Sakimura, "JSON Web Token (JWT)", RFC 7519, DOI 10.17487/RFC7519, , <https://www.rfc-editor.org/rfc/rfc7519>.
- [RFC7540]
- Belshe, M., Peon, R., and M. Thomson, Ed., "Hypertext Transfer Protocol Version 2 (HTTP/2)", RFC 7540, DOI 10.17487/RFC7540, , <https://www.rfc-editor.org/rfc/rfc7540>.
- [RFC7595]
- Thaler, D., Ed., Hansen, T., and T. Hardie, "Guidelines and Registration Procedures for URI Schemes", BCP 35, RFC 7595, DOI 10.17487/RFC7595, , <https://www.rfc-editor.org/rfc/rfc7595>.
- [RFC9114]
- Bishop, M., Ed., "HTTP/3", RFC 9114, DOI 10.17487/RFC9114, , <https://www.rfc-editor.org/rfc/rfc9114>.
- [RFC9162]
- Laurie, B., Messeri, E., and R. Stradling, "Certificate Transparency Version 2.0", RFC 9162, DOI 10.17487/RFC9162, , <https://www.rfc-editor.org/rfc/rfc9162>.
- [RFC9943]
- "*** BROKEN REFERENCE ***".
Appendix A. Changes from v06
Version 07 confirms IANA registrations completed since v06, formalizes URI grammar and Identity Document terminology to match deployment, and makes architectural commitments that v06 prepared but did not lock.¶
A.1. Substantive Changes
The following substantive changes were made:¶
-
IANA registrations confirmed. The
agtp://URI scheme is registered under IANA per [RFC7595]; port 4480 is registered under unified service namesagtp(TCP/TLS) andagtp-quic(QUIC) per [RFC6335]. The IANA Considerations section (Section 11.1) and Stack Position section now state the registered values rather than "TBD." The pre-publication prohibition language has been removed.¶ -
AGIS deprecated; AGTP-API introduced. The Agentic Grammar and Interface Specification (AGIS) and the previously-proposed Agent Method Grammar (AMG) and AGTP-Methods drafts are deprecated and replaced by a single unified companion specification, AGTP-API [AGTP-API]. AGTP-API consolidates the method catalog, path grammar, endpoint primitive, semantic block, schema validation, server manifest format, per-server method policy (carried as the manifest's
policies.methodssub-block), PROPOSE and synthesis semantics, and structural rejection codes (404, 405, 459, 460) into a single document because they describe a single concept: what makes a valid agent-server contract. TheMethod-Grammarheader is removed; servers validate inbound method names against the AGTP-API approved verb list directly. The 454 Grammar Violation status code from earlier draft language is removed; method violations now return 459 (verb not in approved list) or 460 (path violates path grammar). The grammar-validation pathway is replaced by the AGTP-API contract validation pathway.¶ -
Status code renumbering and new contract-level codes. AGTP-specific status codes have been moved out of HTTP-assigned numeric ranges to avoid semantic collision with HTTP status codes that may appear in carried payloads. The mapping from v06: 451 → 455 (Scope Violation), 452 → 456 (Budget Exceeded), 453 → 457 (Zone Violation), 455 → 458 (Counterparty Unverified). New codes added per the AGTP-API contract model: 261 (Negotiation In Progress), 262 (Authorization Required; consolidates PROPOSE-time authorization, missing scope at endpoint dispatch, wildcards refused, and anonymous-discovery blocked), 263 (Proposal Approved), 405 (Method Not Allowed; method+path policy reject), 459 (Method Violation; method not in AGTP-API catalog), 460 (Endpoint Violation; path violates AGTP-API path grammar), 463 (Proposal Rejected). 408 (Timeout) and 410 (Gone) retain their HTTP code numbers and are registered with AGTP-specific semantics. The 461, 462, 464, 465, and 552-555 ranges are reserved for future AGTP expansion.¶
-
Form 1a URI grammar. A new canonical URI form
agtp://{agent-id}@{host}[:{port}](Form 1a) is introduced for direct addressing where the resolver does not yet have a path from canonical Agent-ID to endpoint. The Agent-ID identifies; the hostname is a resolution hint. Form 3 (path-style domain anchored) is retained as a compatibility form but deprioritized for new deployments.¶ -
Port portion of URIs is OPTIONAL. When the port is omitted from any AGTP URI, clients MUST use the IANA-assigned default port 4480.¶
-
Wire-format framing. A new normative subsection (Section 5.4.1) requires explicit
Content-Lengthframing on every AGTP message and prohibits TLS socket-level half-close (shutdown(SHUT_WR)). This addresses a deployment-level gap surfaced by early implementations.¶ -
Agent Manifest Document renamed to Agent Identity Document. The artifact now matches the IANA-registered media type
application/vnd.agtp.identity+json. The schema is enumerated in §5.5 with field-by-field semantics: 16 REQUIRED fields, 14 RECOMMENDED fields, and 2 CONDITIONAL fields. The previous v06 schema is preserved in substance; the rename and the explicit field-level enumeration are the changes. The signature field is relocated to an envelope specified in [AGTP-CERT].¶ -
Twelve-method protocol floor. The core method set has been redesigned as a twelve-method floor organized into six cognitive verbs (QUERY, DISCOVER, DESCRIBE, SUMMARIZE, PLAN, PROPOSE) and six mechanics verbs (EXECUTE, DELEGATE, ESCALATE, CONFIRM, SUSPEND, NOTIFY). DISCOVER, PLAN, and EXECUTE are new core methods in v07. The v06 core methods BOOK, SCHEDULE, LEARN, and COLLABORATE are demoted to Tier 2 standard extended methods and are now catalogued in the AGTP-API method catalog ([AGTP-API]) rather than this document. EXECUTE absorbs the role earlier discussed for an INVOKE method: it is the generic carrier for application-layer payloads when higher-level frameworks are composed over AGTP.¶
-
Composition with higher-level frameworks promoted to body text. A new normative section (Section 8) establishes AGTP as a substrate for MCP, A2A, ACP, and ANP carried as content types inside AGTP method invocations. The substrate model, EXECUTE-as-carrier rule, precedence rule between AGTP headers and framework payloads, and canonical mapping table are normative in the body. The pre-existing AGMP composition appendix supplements the body section with worked wire examples.¶
-
Trust score scoping. The
trust_scorefield (renamed frombehavioral_trust_score) remains in the Identity Document REQUIRED schema. Computation methodology, freshness requirements, and signature binding are forward-referenced to a new companion specification, AGTP-TRUST [AGTP-TRUST], which will be developed separately.¶ -
.well-known/agtpbootstrap convention. A new subsection under §5.1 specifies that organizations operating an AGTP namespace under a DNS domain SHOULD publish a bootstrap document athttps://{domain}/.well-known/agtpper [RFC8615], declaring the organization's AGTP presence, endpoint, and verification anchors.¶ -
Media type alignment. The wire content type has been updated from
application/agtp+jsontoapplication/vnd.agtp+json(vendor-tree pending standards-tree promotion at RFC publication). The Agent Identity Document usesapplication/vnd.agtp.identity+json. YAML variants are registered for both.¶ -
Agent Genesis taxonomy clarification. The permanent signed governance-layer origin document that establishes an agent's identity is named "Agent Genesis" throughout the specification. The taxonomy is: Agent Genesis (permanent signed origin document) → canonical Agent-ID (256-bit hash, used in all protocol operations) → Agent Certificate (optional X.509 v3 credential for TLS mutual authentication; see [AGTP-CERT]). The cross-layer relationship between Agent Genesis, canonical Agent-ID, and Agent Certificate has been clarified relative to earlier drafts where the origin document carried a different name.¶
-
Runtime Contract Negotiation Substrate (RCNS) framing. A new section (Section 5.3) makes explicit a property that earlier versions implied but did not name: AGTP fixes the protocol surface at twelve methods and negotiates any additional endpoint surface at runtime via PROPOSE governed by AGTP-API, completing in a single round-trip. The RCNS framing is added to the abstract and to Design Principles. The mechanism (PROPOSE method, AGTP-API contract validation, dynamic endpoint negotiation) was present in v06; v07 names the property and exposes it as a defining characteristic of the protocol.¶
A.2. Wire Format Compatibility
The status code renumbering and the rename from "Agent Manifest Document" to "Agent Identity Document" are wire-format-visible changes from v06. Implementations of v06 will require updates to emit and consume v07 codes and document type identifiers. Implementations following v06 may continue to interoperate with v07 servers that operate in a backward-compatibility mode that recognizes both old and new code numbers, but conformant v07 implementations MUST emit only the v07 codes.¶
A.3. Rationale
The v06 specification was prepared concurrently with IANA filings. v07 closes the loop on those filings (URI scheme, port, media types in evaluation) and propagates the architectural decisions that the filings made permanent. The status code renumbering is conservative: keeping 451 collided with RFC 7725 (Unavailable for Legal Reasons), which is an HTTP code that may legitimately appear in payloads carried over AGTP. The collision was harmless in v06 because no production implementations existed; it would not remain harmless as deployment grows.¶
The twelve-method floor reflects roughly two years of accumulated experience with the v00 through v06 method sets. The cognitive / mechanics organization is informational rather than normative, but provides a stable mental model that constrains future additions to the floor and guides extended-method design. EXECUTE absorbing INVOKE removes a naming ambiguity: EXECUTE is the clearer verb for the operation, and avoiding two methods in the carrier role simplifies composition with higher-level frameworks.¶
Appendix C. Example AGTP Wire Formats
The following examples use a human-readable pseudo-wire format with HTTP-style headers followed by a JSON body. The Content-Type for all AGTP message bodies is application/vnd.agtp+json.¶
C.1. QUERY Request and Response
AGTP/1.0 QUERY
Agent-ID: agt-7f3a9c2d
Authority-Scope: documents:query, knowledge:query
Session-ID: sess-a1b2c3d4
Task-ID: task-0042
Content-Type: application/vnd.agtp+json
{
"task_id": "task-0042",
"parameters": {
"intent": "Key arguments against MCP re: HTTP overhead",
"scope": ["documents:research", "knowledge:session"],
"format": "structured",
"confidence_threshold": 0.75
}
}
AGTP/1.0 200 OK
Task-ID: task-0042
Server-ID: srv-knowledge-01
Attribution-Record: [signed attribution token]
Content-Type: application/vnd.agtp+json
{
"status": 200,
"task_id": "task-0042",
"result": {
"results": [{"content": "...", "source": "doc-agtp-research",
"confidence": 0.91}],
"result_count": 1
}
}
¶
C.2. EXECUTE Request and Response (Carried Application Payload)
The following example shows EXECUTE carrying an application-layer booking action. The Authority-Scope on the request governs the action; a scope mismatch would return 455.¶
AGTP/1.0 EXECUTE
Agent-ID: agt-travel-planner
Authority-Scope: booking:*, calendar:book
Session-ID: sess-trip-2026-04
Task-ID: task-0107
Content-Type: application/vnd.agtp+json
{
"method": "EXECUTE",
"task_id": "task-0107",
"parameters": {
"action": "reserve-flight",
"parameters": {
"resource_id": "flight-AA2847",
"principal_id": "usr-chris-hood",
"time_slot": "2026-04-15T08:00:00Z",
"options": {"seat_preference": "aisle", "class": "economy"}
},
"idempotency_key": "trip-2026-04-15-AA2847"
}
}
AGTP/1.0 200 OK
Task-ID: task-0107
Attribution-Record: [signed attribution token]
Content-Type: application/vnd.agtp+json
{
"status": 200,
"task_id": "task-0107",
"result": {
"booking_id": "BK-2026-0107",
"status": "confirmed",
"resource_id": "flight-AA2847",
"confirmation_code": "XQRT7Y"
}
}
¶
C.3. EXECUTE Carrying an MCP Tool Invocation
The following example shows EXECUTE carrying an MCP tool invocation. The carried payload's Content-Type identifies it as MCP; the AGTP server dispatches to its MCP handler. AGTP supplies identity, authority, and attribution at the wire level; MCP supplies tool semantics in the payload.¶
AGTP/1.0 EXECUTE
Agent-ID: agt-7f3a9c2d
Authority-Scope: mcp:tools:execute, knowledge:query
Session-ID: sess-mcp-bridge-01
Task-ID: task-0210
Content-Type: application/vnd.mcp.tools+json
{
"method": "EXECUTE",
"task_id": "task-0210",
"parameters": {
"action": "mcp.tools.call",
"payload_type": "application/vnd.mcp.tools+json",
"payload": {
"tool": "search_documents",
"arguments": {
"query": "AGTP composition with higher-level frameworks",
"limit": 5
}
}
}
}
AGTP/1.0 200 OK
Task-ID: task-0210
Server-ID: srv-mcp-bridge
Attribution-Record: [signed attribution token]
Content-Type: application/vnd.mcp.tools+json
{
"status": 200,
"task_id": "task-0210",
"result": {
"tool_response": {
"matches": [
{"doc_id": "doc-0042", "score": 0.91, "excerpt": "..."}
]
}
}
}
¶
C.4. ESCALATE Request and Response
AGTP/1.0 ESCALATE
Agent-ID: agt-procurement-03
Authority-Scope: booking:*, payments:confirm
Session-ID: sess-procurement-q2
Task-ID: task-0881
Content-Type: application/vnd.agtp+json
{
"method": "ESCALATE",
"task_id": "task-0881",
"parameters": {
"task_id": "task-0880",
"reason": "scope_limit",
"context": {
"attempted_action": "EXECUTE",
"resource": "vendor-contract-750k",
"block_reason": "Exceeds agent authorization threshold"
},
"recipient": "usr-cfo",
"deadline": "2026-03-19T09:00:00Z"
}
}
AGTP/1.0 202 Accepted
Task-ID: task-0881
Server-ID: srv-escalation-handler
Content-Type: application/vnd.agtp+json
{
"status": 202,
"task_id": "task-0881",
"result": {
"escalation_id": "ESC-0881",
"routed_to": "usr-cfo",
"status": "pending_review",
"task_paused": true,
"estimated_review_by": "2026-03-19T09:00:00Z"
}
}
¶
Appendix D. Comparison Table
| Criterion | AGTP | HTTP/REST | gRPC | AGMP (MCP, A2A, ...) |
|---|---|---|---|---|
| Intent-native methods | Yes (12 Tier 1) | No | No | Partial |
| Intent semantics at protocol level | Native | None | None | Messaging layer only |
| Built-in agent identity | Yes | No | No | No |
| Authority scope enforcement | Protocol-level | None | None | Application-layer |
| Built-in attribution/audit | Yes | No | No | Varies by impl. |
| Transport flexibility | TCP/UDP/QUIC | TCP/TLS | HTTP/2 | HTTP |
| Escalation as first-class primitive | Yes | No | No | No |
| Ecosystem maturity | Proposed | Mature | Mature | Emerging |
| Governance/observability | Native | Manual/bolt-on | Manual | Limited |
| Method registry extensibility | Yes (Expert Review) | Frozen (IETF Review) | N/A | N/A |
| Open core / royalty-free | Yes | Yes | Yes | Yes |
| Agent Identity Document | Native (.agtp format) | None | None | None |
| Tamper-proof identity surface | Yes (hash + signature) | No | No | No |
| Browser-accessible agent identity | Yes (read-only) | No | No | No |
| URI collision prevention | Domain-anchored | N/A | N/A | N/A |
| Agent Genesis | Yes (genesis record) | No | No | No |
| Domain-expiry lifecycle handling | Specified | N/A | N/A | N/A |
| Capability discovery | Native (DESCRIBE) | None | Reflection (partial) | None |
| Resource budget enforcement | Reserved for v01+ (Budget-Limit, 456) | None | None | None |
| Execution attestation (RATS) | Optional (RFC 9334) | None | None | None |
| Observability hooks | Reserved for v01+ (Telemetry-Export) | None | None | None |
| Network zone enforcement | Reserved for v01+ (AGTP-Zone-ID, 457) | None | None | None |
| Session suspension/recovery | Native (SUSPEND method) | None | None | None |
| AGMP composition profiles | Normative body section | N/A | N/A | N/A |
HTTP's method registry (registered with IETF Review per [RFC9110]) is effectively frozen for new semantic methods because any new HTTP method must be backward-compatible with existing HTTP infrastructure globally. AGTP's Expert Review + published spec procedure enables the protocol to evolve its method vocabulary as the agent ecosystem develops, without the backward-compatibility constraints of the HTTP method space.¶
Appendix E. Glossary
- Agent:
-
A software system that executes tasks, makes decisions, and takes actions without continuous human supervision per transaction.¶
- AGMP (Agent Group Messaging Protocol):
-
The collective term for higher-layer AI agent messaging standards that operate over AGTP as their transport substrate, including MCP, A2A, ACP, and ANP. AGMPs define what agents say. AGTP defines how those messages move. See Section 1.6.¶
- Agent Genesis:
-
The permanent, cryptographically signed origin document issued to an agent at registration time by a governance platform. The source from which the canonical Agent-ID is derived (256-bit hash). Issued once, permanently bound, never reissued. See Section 6.7.¶
- Agent Transfer Protocol (AGTP):
-
The application-layer protocol defined in this document, providing a dedicated transport environment for agent traffic.¶
- Agent-ID:
-
A unique identifier for a specific agent instance. Carried in the
Agent-IDrequest header on non-anonymous AGTP requests, and in theagent_idfield of the Agent Identity Document. In the base spec, derived from the Agent Genesis hash. With [AGTP-CERT], cryptographically bound to a verified identity.¶ - Agent Identity Document:
-
A signed
application/vnd.agtp.identity+jsondocument returned when anagtp://URI is resolved. Derived from the agent's.agentor.nomopackage. Contains identity, lifecycle state, trust tier, trust score, behavioral scope, methods, capabilities, and verification anchors. Never contains executable content. Schema enumerated in Section 5.5.¶ - AGTP-Zone-ID:
-
A network-zone boundary identifier declaring the scope within which a request must be processed. Reserved for v01+: not normatively specified as a request header in this revision. When the header is promoted to normative status, SEPs MUST enforce zone boundaries and return 457 Zone Violation if a DELEGATE request would route outside the declared zone.¶
- Attribution Record:
-
A signed, logged record of an agent action, sufficient for audit and compliance purposes. MAY include RATS attestation evidence per [RFC9334] for hardware-rooted execution proof in high-stakes domains.¶
- Authority-Scope:
-
A declared set of permissions defining what actions an agent is authorized to take, expressed as comma-separated
domain:actiontokens per [RFC9110] list-valued header conventions. See Appendix B for the normative ABNF.¶ - Budget-Limit:
-
A request header declaring the maximum resource consumption the principal authorizes for a method invocation, expressed as comma-separated
unit=valuetokens from the IANA AGTP Budget Unit Registry. Exceeding the declared limit causes 456 Budget Exceeded. Reserved for v01+ per Section 5.5.¶ - Delegation Chain:
-
An ordered record of Agent-IDs representing the sequence of delegations that led to the current request.¶
- DESCRIBE:
-
An AGTP cognitive method returning the declared capabilities, supported modalities, method vocabulary, and versioned feature set of a specific agent endpoint. Used for pre-task negotiation.¶
- DISCOVER:
-
An AGTP cognitive method returning a list of candidate Agent-IDs matching specified criteria. Distinguished from QUERY: DISCOVER returns endpoints to talk to; QUERY returns data to consume.¶
- EXECUTE:
-
An AGTP mechanics method that invokes a specific action with parameters or carries an application-layer payload from a higher-level framework (MCP, A2A, ACP). The generic carrier method for composition.¶
- PLAN:
-
An AGTP cognitive method that produces an unexecuted sequence of actions toward a stated goal. The result is a plan the requesting agent or its principal reviews before committing to execution.¶
- AGTP-API:
-
The IETF companion specification [AGTP-API] that defines the contract layer for AGTP: the curated method catalog, path grammar, endpoint primitive, semantic block, schema validation, server manifest format, per-server method policy (the
policies.methodssub-block of the manifest), and PROPOSE and synthesis semantics. AGTP-API is the authoritative source for what makes a valid agent-server contract over AGTP. AGTP-API supersedes the earlier AGIS draft and the proposed AGTP-Methods draft, consolidating their concerns into a single specification.¶ - Contract:
-
The protocol-level definition of what an AGTP interaction means: the verb (drawn from the AGTP-API method catalog), the path (conforming to AGTP-API path grammar), the semantic block, the input and output schemas, the authority requirements, and the composition rules that govern synthesis. Distinguished from middleware concerns ("plumbing"): rate limiting, observability, caching, custom authentication, and request transformation. The protocol governs contracts; middleware governs operations.¶
- Identity-first architecture:
-
The architectural commitment that the canonical Agent-ID (256-bit content-addressed identifier) is the authoritative identity primitive in AGTP. Hosting, DNS anchors, and other resolution paths are aliases.¶
- ESCALATE:
-
An AGTP method representing an agent's intentional deferral of a decision or action to a human principal or higher-authority agent. A first-class method, not a failure code.¶
- Governance Token:
-
A signed, time-limited JWT artifact encoding a specific governance verdict for a specific action. The runtime companion to the Agent Genesis. Default TTL: 30 seconds. Must not be reused.¶
- Intent Verb:
-
An AGTP method name expressing the agent's purpose, as distinguished from HTTP resource-operation verbs (GET, POST, PUT, DELETE).¶
- Method Registry:
-
The IANA-maintained registry of valid AGTP method names and their specifications. Registration requires Expert Review and a published specification.¶
- Principal:
-
The human, organization, or system that authorized an agent to act and is accountable for its actions.¶
- Principal-ID:
-
The identifier of the principal on whose behalf an agent operates. Carried in the agent identity document referenced by
Agent-ID; not transmitted as a separate request header.¶ - Scope-Enforcement Point (SEP):
-
An AGTP-aware infrastructure component, load balancer, gateway, proxy, that enforces Authority-Scope and AGTP-Zone-ID compliance on AGTP requests without application-layer access. Requires [AGTP-CERT].¶
- Scope Violation (455):
-
An AGTP status code returned when an agent requests an action outside its declared Authority-Scope. A governance signal, not a protocol error. MUST be logged.¶
- Session:
-
An AGTP persistent connection context shared across multiple method invocations within a single agent workflow.¶
- SUSPEND (method):
-
An AGTP Tier 1 core method that places a specific active session into a recoverable paused state, issuing a single-use base64url-encoded 128-bit resumption nonce. Session-scoped; does not affect registry lifecycle state. Category: ORCHESTRATE.¶
- Trust Tier:
-
A classification (1, 2, or 3) assigned to an agent at registration based on the strength of identity verification. Tier 1 requires one of three verification paths (DNS-anchored, log-anchored, or hybrid) and a
.nomogoverned package. Tier 2 is org-asserted without cryptographic verification. Tier 3 is experimental, not production-eligible.¶ - 551 Authority Chain Broken:
-
An AGTP status code returned when one or more entries in the Delegation-Chain header cannot be verified as part of a valid and continuous delegation sequence. MUST be logged.¶
Appendix F. AGTP Composition with AGMPs
This appendix supplements Section 8 with worked wire examples for AGMP messages (MCP, A2A, ACP) carried over AGTP. The strategic positioning, substrate model, precedence rules, and canonical mapping table are normative in Section 8. Full composition specifications are provided in [AGTP-COMPOSITION].¶
F.1. Wire Example: A2A Task over AGTP
The following example shows an A2A task carried over AGTP DELEGATE. A2A task identity, message, and artifacts ride in the body; AGTP identity, authority, delegation chain, and attribution ride at the wire level.¶
AGTP/1.0 DELEGATE
Agent-ID: agtp://agtp.acme.tld/agents/orchestrator
Authority-Scope: agents:delegate, documents:query
Delegation-Chain: agtp://agtp.acme.tld/agents/orchestrator
Session-ID: sess-a1b2c3d4
Task-ID: task-0099
Content-Type: application/vnd.agtp+json
{
"method": "DELEGATE",
"task_id": "task-0099",
"parameters": {
"target_agent_id": "agtp://agtp.acme.tld/agents/analyst",
"authority_scope": "documents:query",
"delegation_token": "[signed token]",
"task": {
"a2a_task_id": "a2a-task-7f3a",
"message": "Summarize Q1 financial reports",
"artifacts": []
}
}
}
¶
F.2. Wire Example: MCP Resource Fetch over AGTP
The following example shows an MCP resource fetch carried over AGTP QUERY. The fetch is read-only and naturally maps onto QUERY (rather than EXECUTE) because no application-side action is performed.¶
AGTP/1.0 QUERY
Agent-ID: agtp://agtp.acme.tld/agents/assistant
Authority-Scope: documents:query, knowledge:query
Session-ID: sess-mcp-b2c3d4
Task-ID: task-0100
Content-Type: application/vnd.agtp+json
{
"method": "QUERY",
"task_id": "task-0100",
"parameters": {
"intent": "fetch document corpus for Q1 financial analysis",
"scope": ["documents:financial"],
"modality": "mcp.resource",
"mcp_resource_uri": "mcp://corpus/financial/q1-2026"
}
}
¶
For an MCP tool-call example using EXECUTE as the dispatch method, see the wire-format examples in Appendix C.¶