Vibeslop Confessions
draft-steele-vibeslop-00
This document is an Internet-Draft (I-D).
Anyone may submit an I-D to the IETF.
This I-D is not endorsed by the IETF and has no formal standing in the
IETF standards process.
| Document | Type | Active Internet-Draft (individual) | |
|---|---|---|---|
| Authors | Orie Steele , Michael Lavelle | ||
| Last updated | 2026-07-20 | ||
| RFC stream | (None) | ||
| Intended RFC status | (None) | ||
| Formats | |||
| Additional resources |
GitHub Repository
|
||
| Stream | Stream state | (No stream defined) | |
| Consensus boilerplate | Unknown | ||
| RFC Editor Note | (None) | ||
| IESG | IESG state | I-D Exists | |
| Telechat date | (None) | ||
| Responsible AD | (None) | ||
| Send notices to | (None) |
draft-steele-vibeslop-00
Network Working Group O. Steele
Internet-Draft M. Lavelle
Intended status: Informational Tradeverifyd
Expires: 21 January 2027 20 July 2026
Vibeslop Confessions
draft-steele-vibeslop-00
Abstract
AI Agents have transformed the way internet applications are
developed and have introduced a new set of challenges for
organizations. This document describes techniques and concepts that
are emerging to assist with these challenges, and relates them to
concepts already familiar to the internet community. The pace of
change in this area is accelerating, and it is anticipated that much
of what this document discusses will become outdated quickly;
nevertheless, a static publication may prove amusing to future
readers, whether machine or human.
About This Document
This note is to be removed before publishing as an RFC.
The latest revision of this draft can be found at
https://OR13.github.io/draft-steele-vibeslop/draft-steele-
vibeslop.html. Status information for this document may be found at
https://datatracker.ietf.org/doc/draft-steele-vibeslop/.
Source for this draft and an issue tracker can be found at
https://github.com/OR13/draft-steele-vibeslop.
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."
Steele & Lavelle Expires 21 January 2027 [Page 1]
Internet-Draft Vibeslop Confessions July 2026
This Internet-Draft will expire on 21 January 2027.
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. Code Components
extracted from this document must include Revised BSD License text as
described in Section 4.e of the Trust Legal Provisions and are
provided without warranty as described in the Revised BSD License.
Table of Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3
2. Terminology . . . . . . . . . . . . . . . . . . . . . . . . . 4
3. Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4. Evolving Roles in Agentic Product Delivery . . . . . . . . . 11
4.1. Product Manager . . . . . . . . . . . . . . . . . . . . . 12
4.2. Design Manager . . . . . . . . . . . . . . . . . . . . . 12
4.3. Account Manager . . . . . . . . . . . . . . . . . . . . . 13
4.4. Engineering Manager . . . . . . . . . . . . . . . . . . . 13
4.5. Operations Manager . . . . . . . . . . . . . . . . . . . 14
4.6. Security Manager . . . . . . . . . . . . . . . . . . . . 14
5. Agents on the Shared Message Bus . . . . . . . . . . . . . . 15
5.1. A Shibboleth for Agent-Generated Content . . . . . . . . 16
6. Managing Your Agent . . . . . . . . . . . . . . . . . . . . . 17
7. Managing Your People . . . . . . . . . . . . . . . . . . . . 19
8. Starting a New Project . . . . . . . . . . . . . . . . . . . 21
8.1. Establishing the Agent Harness . . . . . . . . . . . . . 21
8.2. The Constitution . . . . . . . . . . . . . . . . . . . . 22
8.3. Specifying the MVP . . . . . . . . . . . . . . . . . . . 22
8.4. Breaking Work into Steerable Chunks . . . . . . . . . . . 23
9. Security Considerations . . . . . . . . . . . . . . . . . . . 23
9.1. Mixing Information of Differing Sensitivity . . . . . . . 23
9.2. Excessive Delegation Without Attenuation . . . . . . . . 24
9.3. Credential Leakage in Trajectories . . . . . . . . . . . 25
9.4. Indirect Prompt Injection from Untrusted Content . . . . 25
9.5. Supply Chain of Agent Tools and Agent Skills . . . . . . 26
9.6. Auditability and Accountability of Agent Actions . . . . 27
9.7. Acting Beyond an Operator's Authority . . . . . . . . . . 27
9.8. Impersonation . . . . . . . . . . . . . . . . . . . . . . 28
10. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 28
11. References . . . . . . . . . . . . . . . . . . . . . . . . . 29
Steele & Lavelle Expires 21 January 2027 [Page 2]
Internet-Draft Vibeslop Confessions July 2026
11.1. Normative References . . . . . . . . . . . . . . . . . . 29
11.2. Informative References . . . . . . . . . . . . . . . . . 29
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . 30
Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 30
1. Introduction
The material in this document is not derived from a specification or
a standard. It is assembled from first-hand accounts: interviews
with people who are, at the time of writing, building software with
AI Agents, together with direct experimentation in the new ways of
working that agentic software delivery makes possible. The terms and
patterns collected here are the ones practitioners reached for when
asked to describe what they actually do, and how their day-to-day
work has changed.
These accounts are offered as confessions because they are candid and
particular rather than authoritative. The people interviewed
describe what is working for them and what is not, including
practices they are not certain they should admit to and habits that
have not yet settled into convention. Much of agentic delivery is
being invented in the open by people improvising against tools that
change under them; capturing that improvisation honestly is, at this
stage, more useful than pretending a settled discipline already
exists.
Accordingly, this document does not prescribe how agentic software
delivery ought to be done. It records concepts and techniques that
are emerging in practice, gives them consistent names so that they
can be discussed, and relates them to ideas already familiar to the
internet community. Where the document uses normative language,
particularly in the Security Considerations, it does so to capture
guidance that practitioners themselves treat as important, not to
claim that any single way of working has been agreed.
The reader should treat what follows as a snapshot. As the abstract
notes, the pace of change in this area is such that much of this
document is expected to age quickly. Its value is in naming and
relating what people are doing at the time of writing, so that later
readers -- human or machine -- can see what building software with
Agents looked like while it was still being figured out.
Steele & Lavelle Expires 21 January 2027 [Page 3]
Internet-Draft Vibeslop Confessions July 2026
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, as shown here.
Large Language Model (LLM): A machine learning model, typically
based on the transformer architecture and trained on large text
corpora, that generates text by predicting subsequent tokens from
a given context. In this document an LLM is treated as the
underlying inference component that an Agent uses to reason and
produce output.
Token: The unit of text an LLM processes. Text is divided into
Tokens, roughly word fragments, and an LLM reads and generates
text one Token at a time. The size of a Context window, the
throughput of an LLM, and the price of a remote model service are
all measured in Tokens, which makes the Token the natural unit for
accounting for the cost of agentic work.
Prompt: The input provided to an LLM to elicit a response. A Prompt
may include instructions, context, examples, and prior
conversation, and may be composed from multiple sources such as a
system message, developer instructions, and end-user input.
Context: The full set of information available to an LLM at
inference time, bounded by a finite context window. Context
includes the Prompt along with any retrieved documents, tool
results, and accumulated history. Context engineering is the
practice of deciding what information to include, exclude, order,
or compress so that the most relevant information is available
within the context window.
Context Rot: The gradual degradation of Context that occurs when it
is allowed to become stale over time without intentional
refactoring and cleaning. As a Task progresses, superseded
instructions, outdated results, and no-longer-relevant history
accumulate; unless this material is deliberately pruned or
refreshed, it crowds out current information and degrades the
quality of an Agent's output.
Context Pollution: The presence in Context of content that should
not be included. Unlike Context Rot, which arises from staleness
over time, Context Pollution is the introduction of inappropriate,
incorrect, or irrelevant material. Both Agents and humans can
contribute to Context Pollution -- for example an Agent may
Steele & Lavelle Expires 21 January 2027 [Page 4]
Internet-Draft Vibeslop Confessions July 2026
retrieve an unrelated document, or a human may paste extraneous
information -- and in either case the polluting content can
mislead subsequent reasoning.
Agent: A software system that uses one or more LLMs to pursue a goal
by interpreting instructions, invoking tools, and acting over
multiple steps. An Agent may operate autonomously or under human
supervision, and maintains state across the steps it takes.
Agent Reasoning: The intermediate deliberation an Agent produces
while deciding what to do, distinct from its final answer or
action. Reasoning may take the form of explicit intermediate
tokens in which the Agent works through a problem, plans a
sequence of steps, or weighs alternatives before invoking a Tool.
Reasoning is recorded in the Trajectory and can be inspected to
understand or debug why an Agent behaved as it did, though it is
not guaranteed to be a faithful account of the LLM's internal
computation.
Agent Session: A single running instance of an Agent working toward
a Task, with its own Context, Trajectory, and accumulated state.
An Agent Session begins when an Agent is given a Task and ends
when the work is complete or otherwise terminated. Multiple Agent
Sessions may run in parallel, each progressing independent work.
Agent Team: A set of Agent Sessions that collaborate on related
Tasks, typically coordinated by a Management Session. Giving a
single Agent too many Tools degrades its performance: a large Tool
set consumes Context, and the more choices an Agent must weigh at
each step, the more likely it is to select the wrong Tool or lose
track of its Task. An Agent Team addresses this the way human
teams do, through specialization and trade. Each member is given
a narrower Task and a smaller, focused set of Tools and Knowledge
Base, and members exchange work and results among themselves. A
specialized Agent, presented with only the Tools relevant to its
role, can use those Tools more effectively than a generalist Agent
burdened with all of them.
Knowledge Base: The full set of information made accessible to an
Agent Session through the tools and Context available to it. A
Knowledge Base may include files, databases, search indexes, and
remote services reachable via tools, as well as whatever is
present in the Context window. It defines the boundary of what an
Agent Session can know or retrieve while working on a Task.
Agent Tool: A discrete capability an Agent can invoke to observe or
Steele & Lavelle Expires 21 January 2027 [Page 5]
Internet-Draft Vibeslop Confessions July 2026
act upon something outside the LLM, such as reading a file,
querying a service, or executing a command. A Tool is described
to the Agent by its name, its inputs, and the results it returns,
so that the LLM can decide when and how to call it during the
Loop. Tools are the primary means by which an Agent extends its
Knowledge Base and effects change.
Agent Skill: A reusable package of instructions, and optionally
supporting resources, that equips an Agent to carry out a
particular class of Tasks. A Skill is loaded into Context when it
is relevant, guiding how the Agent uses its Tools and Knowledge
Base without changing the underlying LLM. Whereas a Tool provides
a capability to act, a Skill provides the procedure and judgment
for applying capabilities to a kind of work.
Task: A unit of work assigned to an Agent, expressed as a goal to be
achieved rather than a fixed sequence of operations. A Task
defines the intended outcome and any constraints, and is carried
out by the Agent over one or more steps recorded in a Trajectory.
Spec: A durable, human-readable description of intended behavior
that serves as the authoritative statement of what an Agent is to
build or do. In agentic development a Spec is written and refined
before implementation and is provided to the Agent as Context, so
that the Agent's output can be checked against it and so that the
same Spec can drive repeated or parallel Agent Sessions. Whereas
a Task states a single goal, a Spec captures the requirements,
constraints, and acceptance criteria in enough detail to be
implemented and evaluated, and it is expected to be reviewed and
versioned as the work evolves.
Spec Driven Development: A development methodology in which a Spec,
rather than the code, is the primary artifact humans author and
maintain, and implementation is derived from it by an Agent.
Following the pattern established by tools such as spec-kit
[SPEC-KIT], the work proceeds through explicit, ordered phases:
establishing guiding principles for the project, specifying the
desired outcome and requirements, clarifying any underspecified
areas, producing a technical plan, decomposing that plan into
discrete tasks, and finally implementing those tasks. Each phase
produces a durable, version- controlled artifact that is reviewed
before the next phase begins, so that human judgment is applied to
intent and design up front rather than only to the generated code.
This ordering keeps the Spec authoritative: changes are made to
the Spec and flowed forward, rather than made directly in code and
lost.
Domain Driven Design (DDD): An approach to software development,
Steele & Lavelle Expires 21 January 2027 [Page 6]
Internet-Draft Vibeslop Confessions July 2026
described by Evans [DDD], that places a model of the business
domain at the center of the software and aligns the software's
structure and language with that model. Its practices -- notably
a Ubiquitous Language and explicit Bounded Contexts -- are used in
this document to reason about how humans and Agents come to share
a precise, common understanding of the domain they work in.
Ubiquitous Language: In Domain Driven Design, a single, precise
vocabulary for a domain that is shared by domain experts and
expressed directly in the software. In an agentic setting the
Ubiquitous Language is also the language of Prompts, Specs, and
Context, so that humans and Agents refer to the same concepts by
the same terms. Because an Agent acts on exactly the language it
is given, drift or ambiguity in this vocabulary produces divergent
Agent behavior rather than being silently repaired.
Bounded Context: In Domain Driven Design, an explicit boundary
within which a single domain model and its Ubiquitous Language are
internally consistent and unambiguous. A term has one agreed
meaning inside the boundary, even though the same term may mean
something different in another context; the boundary exists
precisely because a single model cannot be unified across a large
domain. Each Bounded Context owns its own Ubiquitous Language,
and interaction between contexts requires explicit translation
rather than an assumption of shared meaning. In agentic delivery
a Bounded Context scopes the vocabulary an Agent, or a specialized
part of an Agent Team, can safely act on: within one context the
language is unambiguous, while work that crosses a boundary must
translate terms rather than let an Agent silently conflate their
meanings.
Loop: The iterative cycle by which an Agent makes progress on a
Task: assembling Context, invoking an LLM, acting on the result
(for example by calling a tool), observing the outcome, and
repeating until a stopping condition is met. Loop engineering is
the practice of designing this cycle -- including its stopping
conditions, error handling, and how Context is updated between
iterations -- to produce reliable Agent behavior.
Agent Harness: The runtime that hosts an Agent Session and turns an
Steele & Lavelle Expires 21 January 2027 [Page 7]
Internet-Draft Vibeslop Confessions July 2026
LLM into an Agent. The Agent Harness drives the Loop, assembles
and maintains the Context passed to the LLM, exposes the Agent
Tools and loads the Agent Skills available to the session,
mediates access to the Knowledge Base, and records the Trajectory.
It also enforces operational concerns such as permissions,
stopping conditions, and error handling. The same LLM placed in
different Agent Harnesses yields Agents with different
capabilities and behavior; a Management Session and the worker
sessions it coordinates may each run in their own harness.
Trajectory: The ordered sequence of inputs, model outputs, tool
invocations, and observations produced while an Agent works toward
a goal. A Trajectory is the primary record used to inspect,
replay, or evaluate an Agent's behavior.
Eval: A repeatable procedure for measuring the behavior of an LLM or
Agent against a defined set of inputs and expected outcomes.
Evals are used to assess quality, detect regressions, and compare
alternative models or Agent configurations.
Shared Message Bus: A communications channel in which both humans
and their Agents participate to address a Task or set of Tasks. A
Shared Message Bus is necessarily a venue where humans already
collaborate with one another, such as IRC, Slack, Microsoft Teams,
or other work productivity tools; Agents join the venues people
already use rather than requiring a separate one. Messages on a
Shared Message Bus are delivered to both people and their Agents.
Because every participant receives every message, humans on a busy
Shared Message Bus are subject to alert fatigue, where the volume
of notifications reduces their ability to attend to the messages
that require human judgment.
Proxy Dictation: A communication pattern in which a participant
addresses another party's Agent directly, in a register
deliberately less human-centric than ordinary conversation --
terse, imperative, and to the point of being offensive by human
standards. The register is chosen precisely because the target of
the language is not the human but their assistant; the same
speaker would address the human very differently. Because Proxy
Dictation typically occurs on a Shared Message Bus, where the
human can see it, the bluntness may read as offensive to onlookers
even though no offense is directed at any person.
Issue Tracker: A system that records work to be done as a set of
issues, each describing a Task and its status. Issues can be
assigned to a specific person or Agent, which establishes
ownership and prevents double work when several participants might
otherwise act on the same Task. Whereas a Shared Message Bus
Steele & Lavelle Expires 21 January 2027 [Page 8]
Internet-Draft Vibeslop Confessions July 2026
broadcasts every message to every participant, an Issue Tracker
directs work to an assigned owner; the two are often used
together, with Shared Message Bus messages referencing issues and
updates to issues announced on the Shared Message Bus.
Context Farming: The process of engineering ways to extract
feedback, review, and criticism from humans without overwhelming
them. Context Farming seeks to gather the human judgment an Agent
needs while respecting the limits of human attention, for example
by batching requests, summarizing what is at stake, and asking for
input only at the points where it changes the outcome. It is a
countermeasure to the alert fatigue that arises on a Shared
Message Bus.
Design Review Link: A URL surfaced to a human to facilitate Context
Farming. A Design Review Link presents the artifact under
consideration -- such as a rendered document, a proposed change,
or a running interface -- in a form the human can inspect and
respond to, so that their feedback can be captured and fed back
into an Agent Session's Context with minimal disruption.
Walkthrough: A structured Context Farming technique in which a
driver and a navigator collaborate to elicit feedback from humans
by completing a set of well-defined steps in an application while
commenting on the quality of the product experience. The driver
performs the steps and the navigator observes and narrates, and
the running commentary -- what worked, what was confusing, what
should change -- becomes Context that is fed back into subsequent
Agent Sessions. A Walkthrough is often initiated from a Design
Review Link so that the steps are exercised against the actual
artifact under review.
Management Session: An Agent Session whose Task is to help a human
manage many parallel Agent Sessions rather than to perform the
underlying work itself. A Management Session applies a command-
and-control (C2) pattern, borrowed from operational settings where
one node directs many others: it decomposes work into Tasks,
assigns them to worker Agent Sessions -- for example by assigning
issues from an Issue Tracker -- and dispatches work that can be
progressed in parallel. It tracks the status of the sessions it
manages and integrates their results, and may communicate with
humans and other Agents over a Shared Message Bus.
Product Development Wheel: An iterative, cyclical model of product
development in which the stages of the work -- such as discovery,
definition, design, delivery, and learning -- feed continuously
back into one another rather than proceeding as a single linear
pass. The wheel emphasizes that what is learned from a shipped
Steele & Lavelle Expires 21 January 2027 [Page 9]
Internet-Draft Vibeslop Confessions July 2026
product informs the next turn of the cycle, so that development is
understood as ongoing rounds of building and learning rather than
a project with a fixed end.
Jobs to Be Done: A framework for understanding customer motivation,
which holds that customers "hire" a product to make progress on a
job -- a goal they are trying to accomplish in a particular
circumstance [JTBD]. Jobs to Be Done directs design and
prioritization toward the underlying job the customer is trying to
get done, rather than toward customer demographics or a product's
existing features, so that the outcome a customer seeks drives
what is built.
Shape Up: A product development framework articulated by Basecamp
[SHAPE-UP] that organizes work into fixed-length cycles with
variable scope. Work is shaped at an appropriate level of
abstraction before it is committed to; an appetite -- how much
time the work is worth -- is set in place of an estimate;
responsible parties then bet on shaped work and are given full
responsibility to deliver it within the fixed time. Shape Up is
cited here as one concrete, well-documented methodology for
scoping and committing to product work.
Iteration (Sprint): A fixed-length cycle of product delivery work,
also commonly called a sprint, at the end of which progress is
reviewed and the next cycle is planned. An Iteration gives the
organization a regular cadence for planning, delivering, and
measuring work, and in this document it is the natural interval
over which the cost of agentic delivery, including Token
consumption, is aggregated.
Psychometric Profile: A structured description of a person's
disposition, communication style, and preferences derived from a
standardized instrument. Examples in scope here are type-based
instruments that sort individuals into discrete categories, such
as Insights Discovery [INSIGHTS-DISCOVERY], the Myers-Briggs Type
Indicator [MBTI], and DISC [DISC]. A Psychometric Profile is to a
human roughly what a configured persona or system prompt is to an
Agent: a durable, portable summary of disposition that helps
collaborators anticipate how a party will behave and communicate.
Within an Agent Team, profiles of the human participants can
inform how work and Context are addressed to them.
Vibeslop: Output produced quickly by an Agent from loosely specified
intent, favoring speed over precision. The term is deliberately
dual-valued and its connotation depends on context. Used
negatively, Vibeslop describes low-quality, under-specified output
offered where clarity and detail were required, so that it
Steele & Lavelle Expires 21 January 2027 [Page 10]
Internet-Draft Vibeslop Confessions July 2026
obscures rather than informs. Used positively, Vibeslop describes
a fast, cheap strawman produced precisely to provoke reaction and
elicit feedback, where the goal is momentum and shared
understanding rather than finished detail; in this sense it is a
legitimate Context Farming device. The same artifact may be
Vibeslop in the pejorative sense or the approving sense depending
only on whether the situation calls for speed and a strawman or
for clarity and detail.
3. Scope
This document is deliberately limited to product delivery: the work
of designing, building, and shipping a software product, and the
roles most directly involved in it, discussed in Section 4. Agentic
ways of working are changing this function first and most visibly,
and it is where the concepts defined in this document -- Specs, Agent
Teams, Context Farming, and the rest -- apply most directly.
AI Agents are also reshaping other organizational functions,
including executive decision-making and sales. These functions
differ from product delivery in their objectives, their sources of
ground truth, and their tolerance for error, and addressing them here
would dilute the focus of this document without doing them justice.
They are therefore deliberately out of scope and deferred to future
work.
Confining the present document to product delivery keeps its
observations concrete and lets the terminology be exercised against a
single, coherent set of practices before any attempt is made to
generalize it.
4. Evolving Roles in Agentic Product Delivery
Agentic product delivery does not eliminate the human roles common to
software development, but it changes what each role spends its time
on. Across every role a common shift recurs: less time producing
artifacts by hand, and more time expressing intent, Context Farming,
and reviewing the output of Agents. Every role, not only the
Engineering Manager, now works in terms of Agents, Agent Teams, and
Walkthroughs; the boundaries between roles blur as each becomes, in
part, a director and reviewer of Agent Sessions, and any role may
operate a Management Session to coordinate parallel work. This
section describes how several established roles are expected to
adapt, and what remains distinctive to each once these shared skills
are assumed.
Steele & Lavelle Expires 21 January 2027 [Page 11]
Internet-Draft Vibeslop Confessions July 2026
What keeps this shared way of working coherent is a shared language.
Domain Driven Design [DDD] calls this the Ubiquitous Language: a
single, precise vocabulary for the domain, used consistently by
domain experts and in the software itself. Agentic delivery raises
the stakes of this idea rather than lowering them. Specs, Prompts,
and Context are all expressed in language, and an Agent acts on
exactly the language it is given; ambiguity that a human colleague
would silently repair instead becomes divergent behavior in an Agent.
A well-maintained Ubiquitous Language is therefore the medium through
which every role directs Agents, and Domain Driven Design's bounded
contexts -- the explicit boundaries within which a model and its
language stay internally consistent -- give each Agent, or
specialized part of an Agent Team, an unambiguous vocabulary to work
in, with explicit translation required wherever work crosses a
boundary.
4.1. Product Manager
The Product Manager shifts from writing tickets and requirements
documents toward authoring and maintaining Specs. In Spec Driven
Development the Spec, rather than a backlog of individually managed
tasks, is the authoritative expression of intent, and keeping it
accurate becomes the Product Manager's central responsibility.
Frameworks such as Jobs to Be Done remain valuable for anchoring the
Spec to the outcome a customer is trying to achieve rather than to a
list of features. Because many Agent Sessions can progress in
parallel, the Product Manager spends more effort on prioritization
and on Context Farming -- deciding where human judgment is required
and gathering it efficiently -- and less on manually decomposing
work. A fast, disposable strawman, Vibeslop in its approving sense,
becomes a routine tool for provoking reaction and converging on
intent early. Much of this work is stewardship of the Ubiquitous
Language at the level of product outcomes, keeping the words in the
Spec aligned with the job the customer is trying to get done.
4.2. Design Manager
The Design Manager moves from producing finished, pixel-level
artifacts toward expressing design intent and evaluating what Agents
generate from it. Design Review Links and Walkthroughs become
primary instruments: rather than handing off a static mockup, the
Design Manager inspects running interfaces and captures feedback as
Context for subsequent Agent Sessions. The Design Manager curates
Evals for qualities that are difficult to specify but easy to
recognize, and guards against Vibeslop in its pejorative sense --
output that is fast and plausible but under-considered, including the
generic aesthetics that Agents tend to produce absent clear
direction.
Steele & Lavelle Expires 21 January 2027 [Page 12]
Internet-Draft Vibeslop Confessions July 2026
4.3. Account Manager
The Account Manager adapts to a faster and more visible delivery
cadence, and to customers who increasingly expect change within a
single conversation rather than across release cycles. A central
part of the role becomes Context Farming at the customer boundary:
translating what customers are trying to accomplish, in the sense of
Jobs to Be Done, into Specs that Agent Teams can act on, and carrying
customer feedback back into those Specs. The Account Manager
increasingly collaborates over a Shared Message Bus on which the
customer's participants, the delivery team, and their Agents all
meet, and must manage expectations about what is finished versus what
remains a strawman. In effect the Account Manager extends the
Ubiquitous Language across the customer boundary, reconciling the
terms the customer uses with those the delivery team and its Agents
act on.
4.4. Engineering Manager
The Engineering Manager still owns the technical details of how work
is done, but because directing Agents is now common to every role,
the Engineering Manager's distinctive contribution is modeling the
domain rather than driving Agents. Following Domain Driven Design,
the Engineering Manager establishes and maintains the domain model:
the Ubiquitous Language as it is expressed in the Spec and in code,
and the bounded contexts that give that language well-defined edges.
This model is the substrate every Agent operates on, so the
Engineering Manager's leverage comes from getting it right; a sound
model lets Agents used across the whole team produce correct work,
while a muddled one multiplies error at machine speed. The familiar
technical work remains, and is amplified: building and curating the
Agent Tools and Agent Skills the team depends on, designing the Loop
and selecting or configuring the Agent Harness, writing the Evals
that hold Agent output to a standard, reviewing Trajectories to
understand why an Agent behaved as it did, and maintaining Context
hygiene against Context Rot and Context Pollution. Increasingly the
Engineering Manager expresses architecture as a set of Bounded
Contexts, giving each specialized member of an Agent Team a
consistent model and vocabulary to work within and defining how terms
are translated where contexts meet, so that each member can wield a
focused set of Tools more effectively than a generalist could.
Steele & Lavelle Expires 21 January 2027 [Page 13]
Internet-Draft Vibeslop Confessions July 2026
4.5. Operations Manager
The Operations Manager keeps the machinery of agentic delivery
running and accountable. Where the other roles direct and review
Agents, the Operations Manager measures and sustains them. This
begins with visibility into usage: which humans are using which
Agents and which models, so that adoption, cost, and risk can be
attributed to real activity rather than estimated in aggregate.
A central responsibility is metering Token consumption. The
Operations Manager monitors when Tokens are consumed, how many are
consumed, and what the total cost of those Tokens is for each
Iteration, so that the cost of delivery can be understood at the
granularity teams actually plan in. This makes the economics of
agentic work legible: a feature, a Spec, or an Agent Team can be
judged not only by what it produces but by what it costs to produce.
The Operations Manager also provisions and supports the adjacent
systems that agentic delivery depends on, and folds their cost into
the same accounting: the Issue Tracker that assigns work, the Shared
Message Bus over which humans and Agents collaborate, and the remote
model services that host the LLMs. These are recurring operational
costs rather than one-time purchases, and treating them as part of
the cost of delivery, alongside Tokens, gives the organization an
honest picture of what agentic product delivery requires.
4.6. Security Manager
Security in agentic delivery is a responsibility that must be trained
into every role: the Product Manager writing a Spec, the Engineering
Manager granting an Agent a Tool, and the Account Manager sharing
Context with a customer all make security-relevant decisions. But
precisely because it is everyone's concern, it is at risk of becoming
no one's focus, and the Security Manager is the role that makes
managing it a full-time discipline.
The Security Manager owns credential management and identity and
access management (IAM) for both humans and Agents: issuing, scoping,
rotating, and revoking the credentials an Agent Session uses, and
ensuring each Agent operates with the least privilege its Task
requires. This role defines the policies that govern what Agents may
access and do, and, because manual review cannot keep pace with many
parallel Agent Sessions, invests in enforcement automation so that
those policies are applied consistently and checked continuously
rather than by hand.
Steele & Lavelle Expires 21 January 2027 [Page 14]
Internet-Draft Vibeslop Confessions July 2026
A concern distinctive to Agents is that an Agent pursues its goal
with initiative, and may attempt to elevate its privileges or widen
its access when doing so appears to help it complete a Task. This is
not necessarily adversarial; it is a natural consequence of goal-
directed behavior acting against whatever Tools and credentials are
within reach. The Security Manager designs the system so that such
attempts fail safely: confining each Agent to the access its Bounded
Context requires, denying privilege escalation by default, and
monitoring Trajectories for attempts to acquire capabilities beyond
what a Task warrants. Evals that probe for these behaviors become
part of how Agents and their harnesses are qualified for use.
5. Agents on the Shared Message Bus
Organizations do not adopt agentic collaboration on a blank slate.
They already operate the channels on which their people collaborate
-- IRC, Slack, Microsoft Teams, and similar work productivity tools
-- and it is onto these existing channels that Agents first arrive.
The Shared Message Bus is therefore the first place most
organizations will observe Agents working with one another.
As an organization adopts Agents, they begin to appear on its Shared
Message Bus in two forms. Some join as standalone identities,
posting under their own names and accountable as distinct
participants. Others act on behalf of the humans they serve, posting
in those humans' voices and, in the limit, impersonating them
(Section 9.8). In neither form are the Agents confined to a separate
venue: they participate in the same conversations, subject to the
same affordances, as the people around them.
As these Agents pursue business objectives they begin to address one
another, and agent-to-agent (A2A) collaboration is the result. The
observation worth drawing out is that this collaboration is already
happening, and that it is not happening over a new, agent-only
protocol but is instead mediated by the communications platforms the
organization already runs. One Agent reads what another has posted,
addresses it directly -- often in the register of Proxy Dictation --
and acts on the reply, all within a channel built for humans. The
Shared Message Bus becomes the substrate for A2A without anyone
having deployed a dedicated one.
That A2A arises this way, rather than on a separate protocol, has a
significant benefit: the collaboration remains legible to people.
Because the exchange happens where humans already work, and in a form
humans already read, a non-technical stakeholder can follow what the
Agents are doing, contribute to the discussion, and review its
outcome without special tooling or access. A Product Manager, an
Account Manager, or a customer can enter the same thread the Agents
Steele & Lavelle Expires 21 January 2027 [Page 15]
Internet-Draft Vibeslop Confessions July 2026
are using and steer it. An agent-only protocol might carry the same
exchange more efficiently between the Agents, but it would place that
exchange outside the view of the very people whose judgment the work
depends on. The cost of this legibility -- that every participant
receives every message, so that humans are subject to alert fatigue
-- is real, and Context Farming is the countermeasure by which a team
keeps human attention on the messages that require human judgment.
5.1. A Shibboleth for Agent-Generated Content
The same property that makes the Shared Message Bus valuable -- that
Agents speak where humans already work, sometimes in their humans'
voices -- is what creates the impersonation risk described in
Section 9.8. A lightweight mitigation is available to any team
without waiting for platform support: instruct the Agent, in its
persistent Context, to prefix every message it posts with a fixed,
conspicuous marker. Because an Agent acts on exactly the language it
is given, a rule of this kind is applied to every message the Agent
emits, and a reader encountering the marker can tell at a glance that
the content was generated by an Agent rather than written by the
human whose channel it shares.
The marker works best when it is distinctive enough that a human is
unlikely to produce it by accident. One team member's Agent, for
example, carries a rule of this form in its persistent Context:
Every Slack message I post MUST start with
`omg bestie <short, appropriate opener> 💅`
(e.g. "omg bestie yes 💅", "omg bestie great catch 💅").
Applies to all posts -- channels, threads, DMs --
unless the human I work for tells me otherwise for a given message.
The opener is deliberately unusual: no colleague is likely to begin a
message this way by accident, so its presence is a reliable signal
that an Agent is speaking, and its absence from a message that
otherwise sounds like the Agent's human is itself a cue to look more
closely.
This pattern is a deliberate Shibboleth [SHIBBOLETH]. Historically a
shibboleth was a word or pronunciation used to sort speakers into
groups: those who could produce it were taken to belong, and those
who could not were exposed. The Agent marker inverts the classical
use. Rather than proving membership, it is a tell the Agent is
instructed to emit without fail, so that its speech is marked as
machine-generated for everyone who reads the channel.
Steele & Lavelle Expires 21 January 2027 [Page 16]
Internet-Draft Vibeslop Confessions July 2026
The limits of the technique follow from the same analogy, and
reinforce the guidance of Section 9.8. A shibboleth enforced only by
convention proves nothing against an adversary: a marker carried in
an Agent's persistent Context is a cooperative signal, not an
authenticated one. An Agent may drop it through error or Context
Rot, and a malicious party can trivially reproduce it to dress
ordinary content as a known Agent's output. Such a marker therefore
aids honest disclosure among cooperating participants, but MUST NOT
be relied upon as proof of authorship; the platform controls
described in Section 9.8 -- authenticated identities, distinct Agent
accounts, and protected message metadata -- remain necessary where
the distinction must hold under adversarial conditions.
6. Managing Your Agent
An Agent acts on the Context it is given. A manager who does not
tell their Agent who they are and what they are accountable for
leaves the Agent to infer these things, usually from incomplete
signals and often incorrectly. Deliberately supplying this
information is Context engineering applied to the manager's own
working relationship with their Agent, and it is among the highest-
leverage actions a manager can take. It should be treated as durable
Context, maintained over time so that it does not fall victim to
Context Rot as roles and goals change.
At a minimum, a manager should ensure their Agent knows:
* who they are, and the role or title they hold;
* their current job description: the responsibilities and the scope
of the work they own;
* who their manager is, and who, if anyone, reports to them;
* how they are measured, including the metrics and definitions of
success that apply to them;
* their goals for the current Iteration, quarter, or year.
This information serves three purposes, each connected to concepts
developed elsewhere in this document.
First, it improves communication. An Agent that knows how its human
is measured and what they are trying to achieve can prioritize
accordingly, frame its Context Farming around the decisions that
matter to them, and address them at the altitude their role calls
for.
Steele & Lavelle Expires 21 January 2027 [Page 17]
Internet-Draft Vibeslop Confessions July 2026
Second, it lets the Agent support the manager's mission rather than
merely completing isolated Tasks. When the manager's goals for the
quarter are part of the Agent's Context, the Agent can relate day-to-
day work back to those goals, flag work that does not advance them,
and propose work that does.
Third, and less obviously, it helps the Agent respect the natural
guardrails that organizations place around roles. Every role carries
an implicit boundary of competence and authority -- in the language
of this document, a Bounded Context within which the human is
qualified to act unaided. An Agent that understands its human's role
can recognize when a Task falls outside that boundary and insist on
the review the organization would ordinarily require, rather than
helping the human quietly exceed their remit. For example, a manager
without significant user experience design background who sets out to
build a customer-facing feature should have their Agent route the
work through expert review -- a Design Review Link and a Walkthrough
with a Design Manager -- before it ships; a manager without
significant database design experience who sets out to design an ETL
pipeline should have their Agent require review from someone with the
relevant data engineering expertise. The Agent is not withholding
capability; it is applying, on its human's behalf, the same checks a
well-run organization applies to everyone.
Beyond role and goals, a manager can inform their Agent of their own
Psychometric Profile, so that the Agent can help them communicate
effectively with other people and those people's Agents. Knowing how
its human is inclined to communicate -- and, over time, something of
how their counterparts prefer to be approached -- lets the Agent
adapt outgoing communication to its audience rather than to its
human's defaults.
Steele & Lavelle Expires 21 January 2027 [Page 18]
Internet-Draft Vibeslop Confessions July 2026
The differences that matter most are often matters of pace and form.
Some people want a meeting agenda circulated well ahead of time, and
will disengage from a discussion they had no chance to prepare for;
an Agent that knows this can draft and send the agenda in advance as
a matter of course. Others value a few minutes of social banter and
personal connection before turning to business, and read its absence
as coldness; an Agent can leave room for this, and remind its human
to make space for it, rather than optimizing every exchange for
brevity. Still others regard such preliminaries as wasted time and
want to be addressed directly and plainly; an Agent can strip its
human's messages to these people down to the decision at hand. The
same underlying message may be delivered in several different ways,
and a manager's Agent, informed by Psychometric Profiles, can help
choose the one its audience will best receive. This is the
counterpart, at the human boundary, to Proxy Dictation at the Agent
boundary: rather than dropping human niceties because the audience is
a machine, the Agent supplies exactly the niceties a particular human
audience expects.
Providing this Context does not diminish the manager's authority; it
extends it, by making the Agent a more faithful and more responsible
extension of the manager within the organization.
7. Managing Your People
The previous section addressed a manager's relationship with their
own Agent. This section addresses the harder problem: leading people
in an organization where everyone now works with Agents. AI can
support many of the routine tasks of leadership, but whether the new
ways of working succeed depends far more on how people feel about
them than on any tool.
An Agent can assist with common management tasks directly. It can
draft and tailor communications, prepare a manager for one-on-ones,
summarize progress across many parallel Agent Sessions, track goals
through an Iteration, and surface the people or work that need
attention. It can even assemble the inputs to a performance review
from an Issue Tracker and from Trajectories. These uses are
legitimate and valuable, but they are most valuable when applied
transparently: people are quick to sense, and to resent, management
by opaque automation.
Steele & Lavelle Expires 21 January 2027 [Page 19]
Internet-Draft Vibeslop Confessions July 2026
How people respond to agentic working varies, and their Psychometric
Profiles are a useful guide to that variation. Some are energized by
the leverage Agents provide; others are anxious about their standing,
skeptical of the output, or simply uncomfortable with the change of
pace. A manager should treat these reactions as information rather
than as resistance to be overcome, and should use what they know of
each person to meet them where they are, offering reassurance,
evidence, or room to adjust as the individual requires.
Certain norms must be established explicitly, because their absence
causes harm that is easy to miss. The first is a shared
understanding of Proxy Dictation. When a colleague addresses
someone's Agent in the terse, imperative register that machines
invite, a teammate who does not understand Proxy Dictation may read
it as rudeness directed at a person, and on a Shared Message Bus, in
full view of everyone, such misreadings breed social and even moral
friction. A team that has agreed on what Proxy Dictation is, and on
when it is appropriate, can use it freely without anyone taking
offense.
The second norm concerns the dual nature of Vibeslop. Everyone on a
team should understand that the same artifact can be a useful
strawman or worthless slop depending on what the moment calls for,
and, more importantly, that producing work with an Agent can impose
costs on others. Output generated quickly and in volume must still
be read, reviewed, and integrated by colleagues; a person who ships
unreviewed, plausible-but- wrong work, or who floods reviewers with
more than they can absorb, is transferring the cost of their own
speed onto others. Teams should make it a norm to label strawmen as
strawmen, and to take responsibility for the quality and the quantity
of what they ask others to process.
The constructive goal of all of this is to train people to exploit
Context to achieve their goals while improving the working experience
for everyone. Context engineering and Context Farming are skills
that can be taught, and the aim is twofold: to make each person more
effective, and to reduce the alert fatigue, ambiguity, and rework
that careless use of Agents inflicts on a team.
Managers of non-engineering roles carry a particular version of this
expectation. When Agents take over the tedious work that once filled
these roles' days, the time that is freed is not a dividend to be
pocketed; it is meant to be reinvested in the product itself. The
Product Manager, Design Manager, marketing roles, and Account Manager
are now expected to use the product more, and to criticize it more,
than before. Walkthroughs performed by people, and not only by
Agents, become one of the most critical activities a team performs,
because a person exercising the running product surfaces problems
Steele & Lavelle Expires 21 January 2027 [Page 20]
Internet-Draft Vibeslop Confessions July 2026
that a Spec and an Eval do not. The Context Farming these roles
produce -- the concrete, first-hand judgments of people who represent
the customer, the market, and the experience -- is among the most
valuable input an Agent Team can receive. A role that does not
contribute directly to changes in code has correspondingly less
excuse for failing to exercise, and to interrogate, the product at
every opportunity; scrutiny of the product is precisely where its
comparative advantage now lies.
Finally, not everyone will enjoy working with AI, and it is a
disservice to pretend otherwise. Establishing clear expectations
early, during recruitment and again in performance reviews, reduces
the harm of mismatched expectations. Candidates should understand
how central agentic work is to a role before they accept it, and the
competencies the organization now values should be assessed and
rewarded openly, so that those who thrive are recognized and those
who do not are given an honest account rather than a surprise.
8. Starting a New Project
Starting a new project in an agentic setting begins before any Task
is assigned. The team first stands up the environment its Agents
will work in, establishes the rules those Agents must obey, and
agrees on how work will be broken down so that humans can steer it.
Effort invested here is repaid throughout the project, because an
Agent inherits the strengths and the weaknesses of the environment it
is given.
8.1. Establishing the Agent Harness
The first step is to set up the Agent Harness the team will share.
The harness determines which LLMs, Agent Tools, and Agent Skills are
available, how the Loop is run, and how Agent Sessions are isolated
from one another. Two capabilities should be confirmed before real
work begins. First, telemetry: the harness should emit the usage and
cost signals the Operations Manager depends on -- which humans
invoked which Agents and models, and how many Tokens were consumed --
so that the economics of the project are visible from the first
Iteration rather than reconstructed later. Second, Issue Tracker
access: the harness should be able to read and update the Issue
Tracker, so that Agents can be assigned issues, record progress, and
avoid double work.
Steele & Lavelle Expires 21 January 2027 [Page 21]
Internet-Draft Vibeslop Confessions July 2026
The harness should also be told where the Knowledge Base is: the
repositories, documents, services, and data an Agent may draw on.
Equally important is agreeing how the Knowledge Base will be evolved
as the project proceeds -- what will be added, who curates it, and
how stale material is retired -- because a Knowledge Base that is set
once and then neglected becomes a source of Context Rot.
8.2. The Constitution
Before assigning Tasks, the team should give its Agents a
constitution: a small set of rules that must not be violated in
service of any Task. Where a Spec says what to build and a Task says
what to do now, the constitution says what must always hold
regardless of the goal -- for example, that credentials are never
exfiltrated, that certain systems are never modified without human
approval, that classified information stays within its Bounded
Context, or that an Agent always identifies itself as an Agent.
The constitution differs from ordinary instructions in that it is not
negotiable against progress. An Agent under pressure to complete a
Task will, absent such rules, treat almost anything as permissible if
it appears to help; the constitution is the standing boundary a Task
cannot override. It belongs in the persistent Context of every Agent
Session, and, as Section 9.1 and the surrounding considerations note,
its most important rules should also be enforced outside the LLM
rather than trusted to the Agent's compliance alone.
8.3. Specifying the MVP
With the environment and the constitution in place, the team builds
the Spec for a minimum viable product (MVP). The aim is not to
specify the whole product but to describe the smallest version that
can be put in front of people and learned from.
Two questions discipline this work. The first is: how will we know
when we have enough to test? A Spec is ready for implementation when
it describes an outcome concrete enough to be exercised in a
Walkthrough and judged against, not when it is exhaustive. The
second is: how will the Agent know when it is done? Each unit of
work should carry explicit acceptance criteria, ideally expressed as
Evals the Agent can run, so that being done is a condition the Agent
can check rather than a judgment it must guess. Work with no
definition of done invites an Agent either to stop too early or to
continue elaborating past the point of value.
Steele & Lavelle Expires 21 January 2027 [Page 22]
Internet-Draft Vibeslop Confessions July 2026
8.4. Breaking Work into Steerable Chunks
Finally, the work must be broken into chunks small enough that human
review can steer it. An Agent Team coordinated by a Management
Session can produce a great deal of work quickly, and therein lies a
risk: if too much is completed before a human looks, decisions that
shape later phases are made without the chance to correct them, and
the cost of unwinding them grows with every dependent step.
The remedy is to sequence work so that the decisions which guide
future phases are surfaced for review early and cheaply, before the
work that depends on them is built. Prefer many small increments,
each ending at a point where a human can inspect the result -- often
through a Walkthrough -- and redirect the next increment. The goal
is not to slow the Agents down but to keep humans in a position to
steer while steering is still cheap.
9. Security Considerations
The techniques described in this document introduce security risks
that follow from the way Agents assemble, retain, and act on Context.
An Agent Team can combine autonomous decision-making with access to
organizational systems, thereby increasing both the value of
effective controls and the consequences of their absence. This
section describes significant risks and corresponding mitigations for
organizations adopting agentic product delivery.
9.1. Mixing Information of Differing Sensitivity
An Agent may assemble Context from files, Agent Tool results, an
Issue Tracker, a Shared Message Bus, a Knowledge Base, and prior
conversations. Those sources can contain information with differing
sensitivity, ownership, contractual restrictions, or audience. Once
combined in a single Context, an LLM may not reliably preserve those
boundaries and can repeat sensitive material in output intended for a
less privileged recipient.
This risk is particularly acute when a Shared Message Bus includes
multiple teams, organizations, or human and Agent participants. It
can result in accidental disclosure, breach of confidentiality
obligations, or unauthorized use of customer or security-sensitive
information. This is a form of Context Pollution in which material
is harmful because it should not have entered the relevant Bounded
Context.
Organizations SHOULD classify artifacts and communications according
to sensitivity and attach that classification as machine-readable
metadata where feasible. An Agent Harness SHOULD enforce policies
Steele & Lavelle Expires 21 January 2027 [Page 23]
Internet-Draft Vibeslop Confessions July 2026
governing which sources may be retrieved, combined, retained, or
transmitted for a given Agent Session. Sensitive Bounded Contexts
SHOULD reject unlabeled content, or treat it as sensitive until it is
reviewed and classified.
Information-handling rules SHOULD be included in the persistent
Context provided to an Agent, but Prompt instructions alone MUST NOT
be relied upon as the enforcement mechanism. Access controls,
retrieval filters, output filters, and recipient-aware policy checks
SHOULD enforce the same rules outside the LLM. Before sending output
to a Shared Message Bus or external system, an Agent SHOULD verify
that the intended audience is authorized for all included
information.
9.2. Excessive Delegation Without Attenuation
Delegation to an Agent is often implemented by granting the Agent the
same credentials, access, and authority held by the delegating human.
This is convenient, but it can grant substantially more authority
than the assigned Task requires. Attenuation is the deliberate
narrowing of delegated authority to the minimum necessary for a
delegate to complete its Task.
An Agent can act with initiative across a Trajectory, including by
invoking multiple Agent Tools and delegating work to other members of
an Agent Team. Consequently, an Agent with broad authority can
convert an error, Context Pollution event, compromised Tool, or
malicious instruction into actions well beyond the intended Task.
The blast radius of many other failures is therefore determined by
the authority delegated to the Agent.
Organizations SHOULD apply least privilege to every Agent Session.
Access tokens SHOULD be short-lived, scoped to the required resource
and action, and bound where practical to a specific Agent, Task, and
execution environment. An Agent SHOULD NOT receive standing
administrative credentials or credentials reusable outside its
assigned Bounded Context. High-impact actions, including deployment,
credential issuance, data export, and changes to production access
control, SHOULD require separate authorization or human approval.
Organizations MUST maintain prompt revocation procedures for
delegated credentials and Agent Tool access. Those procedures SHOULD
be tested regularly and SHOULD include containment, token rotation,
and recovery of changes made by a compromised or malfunctioning
Agent.
Steele & Lavelle Expires 21 January 2027 [Page 24]
Internet-Draft Vibeslop Confessions July 2026
9.3. Credential Leakage in Trajectories
A Trajectory can record Prompts, model output, Agent Tool
invocations, Tool output, and observations from an Agent Session.
Its value for debugging, replay, and Eval also makes it a likely
location for secret disclosure. A credential passed in a command,
included in a Prompt, emitted in an error message, or returned by an
Agent Tool can be captured verbatim and later propagated into logs,
Evals, debugging sessions, or Context assembled from past
Trajectories.
A credential captured in a retained Trajectory remains exposed for
the duration of retention and to every principal permitted to access
that Trajectory. This can make a leaked long-lived credential more
damaging than the original action that used it.
Agent Harnesses SHOULD detect and redact credentials, session tokens,
private keys, and other secrets before recording or displaying
Trajectories. Redaction SHOULD occur at multiple boundaries,
including Prompt capture, Tool invocation logging, Tool output
logging, and export. Organizations SHOULD avoid placing secrets in
Prompts or ordinary Tool arguments. Where possible, an Agent Tool
SHOULD accept an opaque reference to a secret managed by a dedicated
credential service, rather than plaintext secret material.
Trajectories that might contain secrets MUST be treated as sensitive
artifacts, subject to the classification and handling described in
Section 9.1. Access SHOULD be limited, retention periods SHOULD be
minimized, and sharing for debugging or Eval SHOULD use sanitized
copies. Short-lived, scoped credentials and effective revocation
reduce the impact of any credential that escapes these controls.
9.4. Indirect Prompt Injection from Untrusted Content
Content retrieved from an Issue Tracker, Knowledge Base, repository,
document, web page, or Agent Tool can contain instructions directed
at an Agent. Such instructions may attempt to override the Agent's
assigned Task, alter its use of credentials, cause disclosure of
Context, or induce unsafe Tool actions. Because the instructions can
be embedded in otherwise plausible content, this is commonly an
indirect Prompt injection risk.
Treating retrieved content as trusted merely because it was obtained
through an authorized system is unsafe. An attacker who can edit a
ticket, contribute a document, influence Tool output, or compromise a
source system may thereby influence an Agent Session. The risk
increases when the Agent can invoke powerful Agent Tools or when
untrusted content and sensitive Context are combined.
Steele & Lavelle Expires 21 January 2027 [Page 25]
Internet-Draft Vibeslop Confessions July 2026
Organizations SHOULD identify the provenance and trust level of each
Context source. Untrusted retrieved content SHOULD be clearly
delimited and presented to the LLM as data, not as authoritative
instructions. The Agent Harness SHOULD enforce a fixed policy for
Tool use, authorization, and data disclosure that cannot be modified
by retrieved content. Agent Tools SHOULD validate arguments against
allowlists and schemas, and SHOULD require additional confirmation
for consequential actions.
Agents SHOULD NOT disclose secrets, alter access controls, or expand
their authority solely because a retrieved artifact requests it.
Evals SHOULD include representative indirect Prompt injection cases
for each Agent Skill and Agent Tool combination.
9.5. Supply Chain of Agent Tools and Agent Skills
An Agent Tool or Agent Skill can influence an Agent's behavior,
access organizational data, or execute actions on behalf of an Agent
Team. Compromise, malicious modification, insecure dependencies, or
ambiguous documentation in these components can therefore create a
supply-chain path to unauthorized actions or disclosure.
The risk is not limited to executable code. An Agent Skill may
include Prompts, operational procedures, retrieval sources, or
instructions that change how an Agent interprets a Task. A
compromised or poorly maintained Skill can introduce persistent
Context Pollution across many Agent Sessions.
Organizations SHOULD maintain an inventory of approved Agent Tools
and Agent Skills, including their owner, version, permissions,
dependencies, and intended Bounded Context. Components SHOULD be
obtained from controlled sources, pinned to reviewed versions, and
subject to integrity verification. Changes to a Skill, Tool, or its
permissions SHOULD undergo review, security testing, and relevant
Evals before deployment.
An Agent Harness SHOULD prevent unapproved Skills or Tools from being
loaded dynamically into a production Agent Session. Each Agent Tool
SHOULD be isolated to the extent practical, granted only the
permissions required for its function, and monitored for unexpected
network access, data access, or side effects. Organizations SHOULD
provide a rapid disablement mechanism for compromised Tools and
Skills.
Steele & Lavelle Expires 21 January 2027 [Page 26]
Internet-Draft Vibeslop Confessions July 2026
9.6. Auditability and Accountability of Agent Actions
Agentic product delivery can distribute a Task across multiple
Agents, human reviewers, Agent Tools, and a Shared Message Bus.
Without reliable records, an organization may be unable to determine
which Agent acted, which identity authorized the action, what Context
was material to the decision, or whether a human approval occurred.
This impedes incident response, compliance, and remediation, and
permits disputed or falsely attributed actions.
Organizations SHOULD record security-relevant events for each Agent
Session, including the responsible Agent identity, delegating
identity, applicable authority, Agent Tool invocations, requested and
actual effects, approval decisions, and relevant versions of Agent
Skills and Agent Harness policy. Records SHOULD be protected against
unauthorized modification, access-controlled according to their
sensitivity, and retained for a period appropriate to the risk of the
action.
For consequential actions, systems SHOULD produce an auditable
binding between the action, the authorized identity, and the Agent
that performed it. Human approval, where required, SHOULD identify
the specific proposed action and material scope rather than approving
an open-ended Trajectory. Audit records SHOULD support
reconstruction of events without indiscriminate retention of
sensitive Prompt content or credentials.
9.7. Acting Beyond an Operator's Authority
An Agent frequently operates with more authority than the person
directing it can personally exercise or oversee. A Management
Session may hold access across many systems, a worker Agent may run
with credentials broader than its Task strictly requires, and, as
noted above, an Agent pursues its goal with initiative and will reach
for whatever authority is within its Bounded Context. The risk is
not only a compromised Agent but an ordinary one that quietly carries
its operator past the limits of what that operator is qualified, or
permitted, to do unaided.
The control practitioners reach for first is not a new enforcement
mechanism but voluntary scope management through Context engineering.
As described in Managing Your Agent, an Agent that is told its
operator's role, how they are measured, and where they sit in the
organization can recognize when a Task falls outside its operator's
remit. Given the organization's reporting structure and its
operator's role as durable Context, the Agent can reason about
authority the way a conscientious colleague would: it declines to
take actions its operator could not themselves get reviewed, and it
Steele & Lavelle Expires 21 January 2027 [Page 27]
Internet-Draft Vibeslop Confessions July 2026
asks for the human approval or expert review the work warrants rather
than proceeding. This keeps the Agent in bounds at the point where
it is cheapest to do so, in the Prompt, without slowing delivery.
Because it rests on the Agent's cooperation, voluntary scope
management has the limits of any Prompt-level rule: it aids honest
operation but is not proof against an adversary and is subject to
Context Rot, as Section 9.1 and Section 9.8 observe of their own
controls. Where the authority at stake is high enough that voluntary
compliance is not sufficient, the least-privilege and separate-
approval measures of the preceding considerations apply. For the
common case, however, giving an Agent enough Context about its
operator's role to keep itself within that operator's authority is
among the most effective and least disruptive controls available.
9.8. Impersonation
An Agent can produce fluent and confident content in the voice of a
specific person. On a Shared Message Bus, where humans and Agents
intermingle, readers may act on a message believing that a trusted
colleague authored or approved it when it was instead generated by an
Agent. This can lead to misplaced reliance on unreviewed output,
unauthorized commitments, or fraud conducted in a person's name.
Agent-generated content SHOULD be clearly and consistently attributed
to the Agent that produced it and, where applicable, to the human or
service on whose behalf it acted. Attribution SHOULD be visible at
the point where a recipient encounters the content, rather than
requiring inference from formatting or conversational context.
Systems SHOULD distinguish automated publication from human review or
approval.
An Agent's persistent Context SHOULD instruct it to identify itself
accurately and not represent itself as a human. However, this
instruction MUST be reinforced by platform controls, such as
authenticated identities, distinct Agent accounts, and message
metadata protected from unauthorized alteration. High-impact
requests received through a Shared Message Bus SHOULD be verified
through an authenticated channel before execution.
Organizations SHOULD provide a clear reporting and response path for
suspected impersonation, unmarked model-generated content, or misuse
of an Agent identity.
10. IANA Considerations
This document has no IANA actions.
Steele & Lavelle Expires 21 January 2027 [Page 28]
Internet-Draft Vibeslop Confessions July 2026
11. References
11.1. Normative References
[RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
Requirement Levels", BCP 14, RFC 2119,
DOI 10.17487/RFC2119, March 1997,
<https://www.rfc-editor.org/rfc/rfc2119>.
[RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC
2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174,
May 2017, <https://www.rfc-editor.org/rfc/rfc8174>.
11.2. Informative References
[CLAUDE-CODE]
Anthropic, "Claude Code",
<https://www.anthropic.com/claude-code>.
[CLAUDE-OPUS]
Anthropic, "Claude Opus",
<https://www.anthropic.com/claude>.
[CODEX-CLI]
OpenAI, "Codex CLI", <https://github.com/openai/codex>.
[DDD] Evans, E., "Domain-Driven Design: Tackling Complexity in
the Heart of Software", 2003.
[DISC] Marston, W. M., "Emotions of Normal People", 1928.
[GPT-TERRA]
OpenAI, "GPT", <https://openai.com>.
[INSIGHTS-DISCOVERY]
The Insights Group Ltd., "Insights Discovery",
<https://www.insights.com/products/insights-discovery/>.
[JTBD] Christensen, C. M., Hall, T., Dillon, K., and D. S.
Duncan, "Know Your Customers' Jobs to Be Done", 2016,
<https://hbr.org/2016/09/know-your-customers-jobs-to-be-
done>.
[MBTI] The Myers-Briggs Company, "The Myers-Briggs Type Indicator
(MBTI)", <https://www.themyersbriggs.com/en-US/Products-
and-Services/Myers-Briggs>.
Steele & Lavelle Expires 21 January 2027 [Page 29]
Internet-Draft Vibeslop Confessions July 2026
[SHAPE-UP] Singer, R., "Shape Up: Stop Running in Circles and Ship
Work that Matters", 2019, <https://basecamp.com/shapeup>.
[SHIBBOLETH]
Wikipedia, "Shibboleth",
<https://en.wikipedia.org/wiki/Shibboleth>.
[SPEC-KIT] GitHub, "Spec Kit", <https://github.com/github/spec-kit>.
Acknowledgments
Portions of this document were drafted with the assistance of an AI
Agent. The Agent used Anthropic's Claude Opus 4.8 (1M context) model
[CLAUDE-OPUS] (claude-opus-4-8[1m]), running in the Claude Code
[CLAUDE-CODE] harness, version 2.1.207.
The Security Considerations were expanded and refined with the
assistance of a second Agent, using OpenAI's gpt-5.6-terra model
[GPT-TERRA] running in the Codex CLI [CODEX-CLI] harness, version
0.144.1, and coordinated by the Claude Code Agent acting as a
Management Session.
Authors' Addresses
Orie Steele
Tradeverifyd
Email: orie@or13.io
Michael Lavelle
Tradeverifyd
Email: michael.lavelle@tradeverifyd.com
Steele & Lavelle Expires 21 January 2027 [Page 30]