AI based Network Management Agent(NMA): Concepts and Architecture
draft-zhao-nmop-network-management-agent-04
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 | XingZhao , Minxue Wang , Bo Wu , Daniele Ceccarelli , Haomian Zheng , Jin Zhou | ||
| Last updated | 2026-02-26 | ||
| RFC stream | (None) | ||
| Intended RFC status | (None) | ||
| Formats | |||
| 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-zhao-nmop-network-management-agent-04
Network Management Operations X. Zhao
Internet-Draft CAICT
Intended status: Informational M. Wang
Expires: 31 August 2026 China Mobile
B. Wu
Huawei
D. Ceccarelli
Cisco
H. Zheng
Huawei
J. Zhou
ZTE
27 February 2026
AI based Network Management Agent(NMA): Concepts and Architecture
draft-zhao-nmop-network-management-agent-04
Abstract
The evolution from Level 3 (assisted automation) to Level 4
(autonomous self-optimization) in Autonomous Networks (AN) introduces
requirements for Agentic capabilities, including intent-based
reasoning, autonomous planning, and context-aware decision-making,
which transcend the static, rule-based logic of traditional SDN
Controllers. This document defines the concept of the Network
Management Agent (NMA), an AI-driven entity designed to embody these
cognitive functions and bridge the gap between service intent and
network operations.
This document also specifies how the NMA utilizes the existing
capabilities of SDN Controllers—such as topology management,
telemetry, and enforcement—to achieve Autonomous L4 without
duplicating policy control functions. It further details the
architectural integration modes and defines the interface
requirements necessary for SDN Controllers to interoperate with NMAs.
Discussion Venues
This note is to be removed before publishing as an RFC.
Discussion of this document takes place on the Network Management
Operations Working Group mailing list (nmop@ietf.org), which is
archived at https://mailarchive.ietf.org/arch/btrse/nmop/.
Source for this draft and an issue tracker can be found at
https://datatracker.ietf.org/doc/draft-zhao-nmop-network-management-
agent/.
Zhao, et al. Expires 31 August 2026 [Page 1]
Internet-Draft Network Management Agent Concept February 2026
Status of This Memo
This note is to be removed before publishing as an RFC.
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 20 April 2026.
Copyright Notice
This note is to be removed before publishing as an RFC.
Copyright (c) 2025 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.
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/.
Zhao, et al. Expires 31 August 2026 [Page 2]
Internet-Draft Network Management Agent Concept February 2026
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 31 August 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. 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 . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1. Motivation: The Gap between AN L3 and L4 . . . . . . . . 4
2. NMA and SDN Controller: Roles and Collaboration . . . . . . . 4
2.1. Why NMA is Required for Autonomous L4 . . . . . . . . . . 5
2.2. Utilizing Existing SDN Controller Capabilities . . . . . 5
3. Terminology . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.1. Acronyms and Abbreviations . . . . . . . . . . . . . . . 6
3.2. Definitions . . . . . . . . . . . . . . . . . . . . . . . 6
4. Reference architecture of NMA and Deployment Modes . . . . . 6
4.1. Intelligent Network Management and Control Framework Based
on NMA . . . . . . . . . . . . . . . . . . . . . . . . . 7
4.2. Deployment modes of NMA . . . . . . . . . . . . . . . . . 9
4.3. Reference Functional Architecture of NMA . . . . . . . . 12
4.3.1. Autonomous Logic Layer . . . . . . . . . . . . . . . 13
4.3.2. Supporting Function Layer . . . . . . . . . . . . . . 14
4.4. Interface Requirements for NMA Integration . . . . . . . 15
5. Operational Agent Example . . . . . . . . . . . . . . . . . . 18
6. Security Considerations . . . . . . . . . . . . . . . . . . . 19
7. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 20
8. Appendix：Definition of L0~L5 levels in Autonomous Network . . 20
9. References . . . . . . . . . . . . . . . . . . . . . . . . . 22
9.1. Normative References . . . . . . . . . . . . . . . . . . 22
9.2. Informative References . . . . . . . . . . . . . . . . . 22
Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 23
Zhao, et al. Expires 31 August 2026 [Page 3]
Internet-Draft Network Management Agent Concept February 2026
1. Introduction
1.1. Motivation: The Gap between AN L3 and L4
The Autonomous Networks (AN) framework [TMF-IG1230] defines a series
of evolution stages from Level 0 (manual) to Level 5 (fully
autonomous) as listed in Appendix I. Current operator networks
typically operate at Level 2 or 3, where automation is primarily
policy-driven and reactive. Achieving Level 4 (L4) requires evolving
from static execution to dynamic assurance.
The initial journey towards L4 could target pragmatic, high-value
scenarios, such as automated Root Cause Analysis (RCA), SLA
assurance, and service restoration. These use cases allow operators
to deploy AI for observability and recommendation, reducing manual
toil while maintaining control.
Traditional SDN Controllers excel at deterministic configuration and
telemetry collection but lack the analytical depth required for these
complex assurance tasks. They execute instructions but cannot
autonomously diagnose the 'why' behind a failure or predict SLA
violations. Introducing AI-driven logic is necessary to bridge this
gap. However, decoupled AI models are insufficient. A new
architectural entity—the Network Management Agent (NMA)—is needed to
integrate AI-based reasoning with SDN control, starting with
assurance use cases and gradually evolving towards full closed-loop
autonomy.
While the key issues after the introduction of AI in network
management include:
1. The application architecture and deployment methods of AI in
network management are still unclear, that is in what form AI can
help network management?
2. The relationship between AI and the existing network controllers
is not clear.
3. New interface capability requirements after AI is introduced are
not clear either.
Therefore, it is necessary to define the general architecture and
application form of AI in network management.
2. NMA and SDN Controller: Roles and Collaboration
Zhao, et al. Expires 31 August 2026 [Page 4]
Internet-Draft Network Management Agent Concept February 2026
2.1. Why NMA is Required for Autonomous L4
Achieving L4 autonomy requires a cognitive loop of Intent
Interpretation, Perception Analysis, and Dynamic Decision-
making—capabilities that extend beyond the native design of
traditional SDN Controllers:
* *Intent Translation (The "Why")*: L4 moves beyond simple API
commands to handling fuzzy, high-level operational intents.
Unlike SDN Controllers, which require precise, low-level technical
parameters (e.g., specific bandwidth values or queue IDs), the NMA
acts as an Agentic Interpreter. It automatically decomposes
abstract goals (e.g., "Ensure optimal experience for VPN users")
into concrete, verifiable technical specifications, handling the
ambiguity and context that traditional controllers cannot resolve.
* *Perception & Contextual Analysis (The "Sense")*: L4 requires
holistic observability not just raw data collection. SDN
Controllers excel at gathering telemetry but lack the ability to
fuse multi-dimensional data (metrics, logs, traces, alarms) to
understand the "state of the network" in a service context. The
NMA combines its own knowledge base and memory, using AI models to
perform Root Cause Analysis (RCA), detect anomalies, and correlate
events across the network to build a comprehensive operational
picture.
* *Autonomous Decision & Policy Synthesis (The "Think")*: L4 demands
the ability to make non-deterministic decisions in response to
unforeseen scenarios. Traditional controllers operate on
deterministic, reactive logic (e.g., "If X, then Y"), which cannot
handle novel failures or complex optimization trade-offs. The NMA
embodies the Decision function, utilizing reasoning capabilities
to synthesize new strategies, weigh potential outcomes, and decide
on the optimal course of action when standard procedures do not
apply, and can be iteratively optimized itself.
Therefore, the NMA serves as the Autonomous Brain (Cognitive Layer)
that defines what needs to happen and why, orchestrating the SDN
Controllers, which act as the Execution function (Control Layer) that
handle how to enforce those decisions on the network infrastructure.
2.2. Utilizing Existing SDN Controller Capabilities
To realize Autonomous L4, the NMA leverages the mature, stable
functions already present in SDN Controllers rather than reinventing
them. NMA is compatible with the YANG-based automation framework
described in [RFC8969], and utilizes the Controller as its primary
execution engine:
Zhao, et al. Expires 31 August 2026 [Page 5]
Internet-Draft Network Management Agent Concept February 2026
* *Model-Based Abstraction*: The NMA interacts with the Controller
through standard YANG Service and Network Models, bridging the gap
between high-level intent and concrete network resources.
* *Telemetry & State Access*: The NMA consumes real-time operational
data and topology information provided by the Controller to
maintain an accurate perception of the network state.
* *Policy Enforcement*: The NMA invokes the Controller's
configuration interfaces to apply changes, relying on the
Controller's built-in validation and transaction capabilities to
ensure stability.
By integrating AI reasoning with this standards-based automation
foundation, the NMA elevates the network from L3 (Automated Control)
to L4 (Autonomous Management).
3. Terminology
3.1. Acronyms and Abbreviations
AI: Artificial Intelligence
LLM: Large Language Model
NMA: Network Management Agent, refers to AI based network management
agent
3.2. Definitions
The document defines the following terms:
*Network Management Agent (NMA):* A network management entity built
based on ML/AI and equipped with the autonomous task processing
capabilities. It can automatically carry out network status
perception, task intent [RFC9315]interpretation, task planning,
decision-making and task execution operations based on user task
intentions or preset goals, so as to achieve closed-loop
processing of scenarios-oriented network management tasks. For
different application scenarios, NMA can be subdivided into
multiple scenario-oriented agents.
4. Reference architecture of NMA and Deployment Modes
Zhao, et al. Expires 31 August 2026 [Page 6]
Internet-Draft Network Management Agent Concept February 2026
4.1. Intelligent Network Management and Control Framework Based on NMA
[RFC8969] proposed the framework for automating service and network
management with YANG. Building on the architecture proposed in
[RFC8969], higher-level intelligent network management and control
can be achieved by adding NMA components. Based on the Figure 3 of
[RFC8969], the layered architecture of intelligent network management
and control after the introduction of NMA is shown in the following
figure. NMA can exist at both the Controller and Orchestrator
levels; for the device layer, due to the constraints on the computing
power of network elements, some end-side AI components may be added
on the device side, while it is unlikely to deploy a complete NMA.
Zhao, et al. Expires 31 August 2026 [Page 7]
Internet-Draft Network Management Agent Concept February 2026
Hierachy NMA interaction
+-------------------------------+
| Orchestrator |
| +---------------------------+ | +-----------+
| | Network Management Agents | | +-|---------+ |
| | (NMAs) | | +-|---------+ |-+
| +---------------------------+ | | NMAs |-+
| +---------------------------+ | +-----^-----+
| | Service Modeling | | |
| +---------------------------+ | |
| +---------------------------+ | | Inter-layer
| | Service Orchestration | | | A2A communication
| +---------------------------+ | |
+-------------------------------+ |
--------------------------------------------------------+--------
+-------------------------------+ |
| Controller | |
| +---------------------------+ | +-v---------+
| | Network Management Agents | | +-|---------+ |
| | (NMAs) | | +-|---------+ |-+
| +---------------------------+ | | NMAs |-+
| +---------------------------+ | +-----------+
| | Network Modeling | |
| +---------------------------+ | NMA1<---------------->NMA2
| +---------------------------+ | Intra-layer
| | Network Orchestration | | A2A communication
| +---------------------------+ |
+-------------------------------+
-----------------------------------------------------------------
+-------------------------------+
| Device |
| +---------------------------+ |
| | End-side AI | |
| +---------------------------+ |
| +---------------------------+ |
| | Device Modeling | |
| +---------------------------+ |
+-------------------------------+
Figure 1: Enhanced intelligent network management and control
framework based on NMA
Zhao, et al. Expires 31 August 2026 [Page 8]
Internet-Draft Network Management Agent Concept February 2026
Among them, there may be interaction requirements between NMAs at
different layers and between different NMAs at the same layer.
Cross-layer NMAs interact through inter-layer Agent-to-Agent (A2A)
communication, while different NMAs within the same layer interact
through intra-layer A2A communication.
This document can be regarded as an enhancement of the intelligent
capabilities of [RFC8969], and subsequent discussions will mainly
focus on the NMAs at the controller layer.
4.2. Deployment modes of NMA
It should be noted that although NMA is depicted inside the
controller in Figure 1, in practice, NMA can also be deployed as an
independent component outside the controller. This document does not
impose mandatory restrictions on the deployment location of NMA. The
two deployment modes can be called: Independent deployment mode and
Integrated deployment mode and are shown in Figure-2, where the NMA
can be part of an existing network controller, or can be an
independent system deployed separately and interacting both with the
controller and the network.
Zhao, et al. Expires 31 August 2026 [Page 9]
Internet-Draft Network Management Agent Concept February 2026
^
|
Extended NBI(including A2U)
|
+-----------------------------v------------------------------+
| Network Controller |
| |
| +--------------------+ +--------------------+ |
| | Original Function <----A2C----> Network management | |
| | Modules | Interface | Agent(NMA) | |
| +--------------------+ +--------------------+ |
| |
+------------------------------^-----------------------------+
|
Extended SBI(including A2N interface)
|
+------------------------------v-----------------------------+
| Physical Network |
+------------------------------------------------------------+
(a) Integrated Mode
^ ^
| |
Northbound Interface(NBI) Agent-to-User Interface(A2U)
| |
+--------------v------------+ +----------v---------+
| | | |
| Network <----A2C----> Network Management |
| Controller | Interface | Agent(NMA) |
| | | |
+--------------^------------+ +----------^---------+
| |
Southbound Interface(SBI) Agent-to-Network Interface(A2N)
| |
+--------------v-----------------------------------v---------+
| Physical Network |
+------------------------------------------------------------+
(b) Independent Mode
Figure 2: Deployment mode of network management agent (NMA)
*Integrated deployment mode:* As shown in Figure-2 (a), NMA is
integrated and deployed with the original network controller, and
the NMA serves as a function of the controller. NMA interacts
with original function modules through internal A2C interface.
The enhanced controller interacts with the underlay physical
network through extended SBI satisfying the A2N interaction
Zhao, et al. Expires 31 August 2026 [Page 10]
Internet-Draft Network Management Agent Concept February 2026
requirements. The specific functional requirements and
information model definition of interfaces mentioned above will be
discussed in Section 4.4.
Integrated mode is targeted at network scenarios with single-
vendor SDN infrastructure and high requirements for service real-
time performance. This mode features deep coupling between the
NMA and the SDN Controller, low decision-making and execution
latency, and simple deployment and operation & maintenance (O&M),
making it suitable for autonomous network management in single-
vendor domains. At the same time, since it is extended on the
basis of an existing SDN controller, the changes and impacts on
the live network are also smaller, which facilitates the
application and evolution of NMA in the live network.
*Independent deployment mode:* As shown in Figure 2 (b), NMA is
independently deployed from the original network controller. NMA
and controller are independent systems. A new east-west interface
needs to be added between the NMA and the controller to achieve
capability calling and result feedback operations. This interface
can be called “Agent-to-Controller Interface”(A2C). In this
deployment mode, controller uses southbound interface (SBI) to
interact with physical network, while an Agent-to-Network
interface (abbreviated as “A2N”) needs to be added between NMA and
the underlying physical network.
Independent mode is applicable to multi-domain, multi-vendor
heterogeneous network environments. Boasting high flexibility and
scalability, this mode enables the NMA to act as a centralized
cognitive brain that orchestrates multiple SDN Controllers to
achieve closed-loop execution of end-to-end service intents.
While the independent deployment mode brings significant flexibility
to the management of large-scale and complex networks, its decoupled
architecture between the NMA and SDN Controllers introduces a series
of potential issues in practical deployment, including management and
O & M conflicts between the two entities, which are mainly reflected
in the following aspects:
*Configuration and policy conflicts:* Concurrent delivery of
configurations to network devices by the NMA and the Controller
may result in configuration conflicts on the devices. In
addition, the NMA generates dynamic control policies based on AI-
driven intent reasoning and real-time network context analysis,
whereas SDN Controllers maintain pre-configured static rule sets
and traditional deterministic automation policies.
Inconsistencies between these two types of policies may lead to
policy execution failures and even service interruptions.
Zhao, et al. Expires 31 August 2026 [Page 11]
Internet-Draft Network Management Agent Concept February 2026
*Inconsistent network state synchronization:* The autonomous
decision-making of the NMA relies on real-time and accurate
network state data (telemetry, alarms, topology) provided by SDN
Controllers. In the independent mode, network transmission
latency and data processing delays between the NMA and Controllers
may compromise the accuracy of the NMA's decision-making.
This document does not mandate a specific deployment mode for the
NMA. When the independent deployment mode is adopted, it is advised
to follow the principle of separation of cognitive decision-making
and execution enforcement: the NMA is responsible for intent
interpretation, context analysis and autonomous decision-making,
while SDN Controllers retain the authorities of policy validation,
resource enforcement and network state management. This ensures the
consistency and effectiveness of the collaborative operation between
the NMA and SDN Controllers.
4.3. Reference Functional Architecture of NMA
In order to achieve above capabilities, by referring to the common AI
agent framework, this document presents the reference functional
architecture of NMA as shown in Figure 3.
+---------------------------------------------------------------------------+
| NMA(Network Management Agent) |
| |
| +-----------------------------------------------------------+ |
| Autonomous | Intent Management | |
| Logic +-----------------------------------------------------------+ |
| Layer +------------+ +------------+ +-----------+ +------------+ |
| | Awareness | | Analysis | | Decision | | Execution | |
| +------------+ +------------+ +-----------+ +------------+ |
|---------------------------------------------------------------------------|
| |
| Supporting +-----------------+ +=================+ +-----------------+ |
| Function | Memory& | | AI Model | | Tool & Function | |
| Layer | Knowledge Base | | Service | | Manager | |
| +-----------------+ +=================+ +-----------------+ |
| |
+---------------------------------------------------------------------------+
Figure 3: Reference function architecture of NMA
Zhao, et al. Expires 31 August 2026 [Page 12]
Internet-Draft Network Management Agent Concept February 2026
The NMA is structured into two primary layers: the Autonomous Logic
Layer, which embodies the autonomous closed-loop from intention to
perception, analysis, decision-making, and execution, and the
Supporting Function Layer, which provides foundational capabilities
to enable autonomous operations.
4.3.1. Autonomous Logic Layer
This layer embodies the intelligent loop of L4 autonomy, translating
service goals into network actions. It mainly includes the following
logical functional modules which are fully consistent with the IAADE-
closed loop of autonomous network defined in TMF (see detailed in
Section 8):
*Intent Management:* This module serves as the entry point for
Intent. It is responsible for receiving high-level goals from
users or orchestration systems, interpreting natural language or
policy objectives, and normalizing them into structured,
verifiable intents that the agent can pursue. It ensures that the
autonomous operations remain aligned with service KPIs. After
interpreting the target intent and reasoning through the necessary
steps to achieve it, this module can orchestrate the sequence of
operations required to progress toward that goal. It breaks down
complex objectives into a sequence of executable sub-tasks (e.g.,
awareness -> analysis -> decision -> execution) and handles
dynamic planning under uncertainty, ensuring that the chosen
course of action aligns with the desired intent.
*Awareness:* This module acts as the intent-driven selective sensing
hub of the NMA, responsible for orchestrating the targeted query
and perception of task-relevant network data. It proactively
initiates data acquisition operations across heterogeneous sources
such as controllers, physical/virtual network devices, etc., with
a core focus on filtering out irrelevant information to collect
only the network data pertinent to the current intent. Covering
critical dimensions including device operational status, link
performance metrics, service traffic statistics, and configuration
parameters, this module lays a precise foundational data base for
the subsequent analysis, decision-making, and execution processes.
*Analysis:* This module serves as the intelligent analytics core,
leveraging the reasoning capabilities of the AI Model Service in
the Supporting Function Layer. It orchestrates advanced
analytical tasks tailored to the specific task intent, including
anomaly detection, root cause analysis (RCA), event correlation,
and impact quantification, etc. By combining real-time perceived
data with historical insights retrieved from the Memory&Knowledge
Base, it transforms raw data into actionable, context-rich network
Zhao, et al. Expires 31 August 2026 [Page 13]
Internet-Draft Network Management Agent Concept February 2026
insights and diagnostic conclusions. It can clearly identify the
root causes of network issues, evaluates the impact of abnormal
states on service objectives, and outputs structured analytical
results that directly guide the strategic decision-making.
*Decision:* This module functions as the strategic decision-making
core of the NMA, responsible for formulating optimal and feasible
operation strategies based on the analytical insights from the
Analysis Management module and the constraints of the original
user intent. It employs AI reasoning capabilities and draws on
the Memory&Knowledge Base to evaluate multiple potential action
paths, selecting the strategy that best aligns with service-level
objectives and network operation rules. It decomposes complex
strategic decisions into a hierarchical, ordered sequence of
executable sub-tasks, defines clear trigger conditions and task
dependencies for each step, and maps these sub-tasks to specific
tools or functions managed by the Tool&Function Manager. This
process ensures that the generated decisions are not only
logically sound but also fully operationalized for subsequent
execution.
*Execution:* This module acts as the intent-closed-loop operational
execution core, tasked with translating the structured sub-tasks
from the Decision Management module into concrete, reliable
network operations. It orchestrates the invocation of appropriate
network interfaces, management tools, and operational functions
via the Tool&Function Manager, executing tasks such as
configuration adjustment, fault remediation, resource scheduling,
and service provisioning in a sequential and controlled manner.
It real-time monitors the execution status of each sub-task,
handles execution exceptions and retries according to pre-defined
rules, and conducts rigorous result validation against the
original user intent and decision criteria. Finally, it feeds
back the execution outcomes, status, and validation results to the
Memory&Knowledge Base and upper-layer modules, forming a complete
closed-loop of autonomous network management driven by intent.
NMA enables the cognitive capabilities on task lifecycle management
procedure described in [RFC8969].
4.3.2. Supporting Function Layer
This layer provides the foundational capabilities and resources
necessary for the Autonomous logic Layer to function effectively.
*Memory & Knowledge Base:* This module serves as the long-term and
Zhao, et al. Expires 31 August 2026 [Page 14]
Internet-Draft Network Management Agent Concept February 2026
short-term memory of the NMA, storing historical operational data,
network topology snapshots, and a comprehensive repository of
expert knowledge including technical documents, troubleshooting
guidelines, and past incident resolution cases, etc. It provides
unified search capabilities across multi-type knowledge sources
such as vector knowledge bases, system online help documentation,
and operation and maintenance data logs. Based on accurate
domain-specific information, this module improves the accuracy and
reliability of NMA’s reasoning and decision-making, enables the
agent to reuse historical experience and expert logic, and ensures
the consistency and effectiveness of autonomous operations.
*AI Model Service:* This module acts as the cognitive engine of the
NMA, providing unified upward exposure of diversified AI
capabilities. It supports not only Large Language Models (LLM)
and other generative AI models, but also classic AI algorithms and
lightweight dedicated models, enabling natural language
understanding, logical inference, time-series analysis and other
intelligent capabilities. It supplies the comprehensive general
and domain-specific intelligence required to drive the core
processes of intent management, perception and analysis, reasoning
and planning, and decision and execution.
It should be noted that the AI Model Service is not limited to
being deployed inside the NMA; it can also be located outside the
NMA, and the NMA can invoke AI model capabilities in real time to
complete relevant reasoning operations.
*Tool & Function Manager:* This module serves as the Gateway to
Reality. It manages the connection between the NMA and external
systems, primarily the SDN Controllers via the A2C (Agent-to-
Controller) interface. It abstracts network functions (e.g.,
configuration, telemetry, simulation, etc.) as invocable "Tools."
This module ensures that the decisions made by the upper layer are
translated into concrete, standard-compliant network operations
(e.g., YANG data manipulation).
4.4. Interface Requirements for NMA Integration
As shown in Figure 2, the interfaces related to NMA include three
types:
1. *Agent-to-User interface (A2U):*the interface between the NMA and
the user, where the user can be upper-layer NMA, controllers or
orchestrators. This interface is used to receive call requests
from users and return task processing results. It should support
both structured and natural language modes. The natural language
interface is mainly used for interaction with humans, while the
Zhao, et al. Expires 31 August 2026 [Page 15]
Internet-Draft Network Management Agent Concept February 2026
structured interface is used for interaction with other upper-
layer systems or other Agents. The Agent-to-Agent (A2A)
interface between NMAs is included in the scope of this
interface. In the independent mode, this interface is a separate
one provided by the NMA to the outside; in the integrated mode,
it is included in the northbound interface of the controller.
Since this interface bridges the NMA with human operators or
higher-level orchestrators. It must support dual-mode
interaction:
*Natural Language Interaction:* For human operators, the
interface must support conversational inputs (e.g., text) and
return structured responses or execution confirmations.
*Structured Intent Interface:* For upper-layer orchestrators or
peer agents, the interface must support structured intent
definitions (e.g., based on YANG models or JSON/GNMI). It
requires:
* Intent Submission: Accepting high-level goals with
constraints (e.g., latency, cost).
* Status Reporting: Providing real-time feedback on intent
fulfillment progress, including intermediate states (e.g.,
"Analyzing", "Planning", "Executing").
2. *Agent-to-Controller interface (A2C):* the interface between NMA
and the controller or the original functional components of the
controller. In the independent mode, this interface is an east-
west interface between the controller and NMA; in the integrated
mode, this interface is an internal interface of the controller
and is not within the scope of this document.
3. *Agent-to-Network (A2N):*the interface between NMA and the
physical network. In the independent mode, this interface is a
southbound interface between the Agent and the network; in the
integrated mode, it is included in the original southbound
interface of the controller.
To elaborate in more detail, when NMAs are deployed in integration
with the controller, as shown in Figure 4, the related interface to
be extended includes:
1. *Extended SBI of the controller:* The southbound interface
between the controller and devices, including the aforementioned
A2N interface function. Theoretically, NMAs will not directly
configure or operate devices; instead, they will call the
Zhao, et al. Expires 31 August 2026 [Page 16]
Internet-Draft Network Management Agent Concept February 2026
original functional modules of the controller for device-related
configuration and management. Therefore, the need for standard
extension of this interface is minimal, and it is not within the
scope of this draft.
2. *Extended NBI:* The northbound interface of the controller. As a
key interface for collaboration between upper and lower layer
systems, this interface needs to realize functions such as
capability discovery and invocation between upper and lower layer
NMAs. Hence, there is a strong demand for its standardization,
and it is necessary to consider the extension of the northbound
interface of the controller oriented to the communication needs
between NMAs. NBI must be augmented to expose the NMA's
cognitive capabilities as Intent-Based RPCs. Unlike standard
configuration RPCs that set specific parameters (e.g., set-
bandwidth), these Intent-Based RPCs accept high-level operational
goals (e.g., optimize-performance or diagnose-incident). This
distinction allows upper-layer systems to invoke autonomous
behaviors that require reasoning and planning—capabilities that
native controller interfaces lack.
In terms of communication channels, the orchestrator and the
controller communicate one-to-one through the northbound
interface. When there is a need for direct communication between
NMAs in the upper-layer orchestrator and those in the lower-layer
controller (A2A Communication), it will manifest as a single
communication channel physically but multiple communication
processes logically (i,e.including multiple A2A communication
processes).
To sum up, entended NBI should handle logical multi-process
multiplexing. Current protocols typically handle a single
request-response session.The extended NBI must support multiple
independent A2A communication processes over a single physical
channel. It must maintain strict context isolation between
different agent tasks (e.g., one diagnosing a fault, another
optimizing QoS) to prevent state interference—a requirement not
addressed in standard HTTP/RPC models.
Besides, there are several internal interfaces within the controller,
which include the interaction interfaces between NMAs within the
controller and the original functional modules of the controller, as
well as the interaction interfaces between multiple NMAs within the
controller. Since all the above are internal implementations of the
controller, there is no need for standardization.
The specific implementation methods, related protocols, etc. of each
interface are to be defined subsequently in other documents.
Zhao, et al. Expires 31 August 2026 [Page 17]
Internet-Draft Network Management Agent Concept February 2026
+-------------------------------------------+
| Orchestrator |
| +-----------+ +-----------+ +-----------+ |
| | NMA1 | | NMA2 | | NMA3 | | ^
| +-----^-----+ +-----------+ +-----^-----+ | |
+-------:-------------^-------------:-------+ |
: | : |
: Logical | Extended : Logical | Extended
: A2A | NBI : A2A | NBI
: | : |
+-------:-------------v-------------:------------------^---+
| : Controller : |
| : : +------------+ |
| +-----v-----+ +-----------+ +-----v-----+ | Original | |
| | NMA1 | | NMA2 | | NMA3 | | function | |
| +-----------+ +-----------+ +-----------+ | modules | |
| +------------+ |
+----------------------------^-----------------------------+
|
| Extented SBI
|
+----------------------------v-----------------------------+
| Physical Network |
+----------------------------------------------------------+
Figure 4: Interfaces to be extended on the controller
5. Operational Agent Example
To address specific operational needs, the NMA architecture supports
multiple specialized agents. These agents function as modular
entities, with the Intelligent Assistant Agent serving as the primary
entry point for interaction, followed by specialized agents such as
Fault management Agent and Optimization Agent:
* *Intelligent Assistant Agent*: Serving as the primary interface
for human operators, this agent leverages LLMs to provide natural
language Q&A and conversational capabilities. It enables users to
perform "one-click" queries for fault descriptions or resource
status. By automatically translating human intent into precise
data retrieval commands, it significantly enhances the efficiency
of knowledge retrieval and daily maintenance support.
* *Network Fault Management Agent*: Focused on service assurance,
this agent leverages comprehensive troubleshooting guides and
expert knowledge bases to support intelligent fault handling. It
implements automated root cause analysis (RCA) and fault impact
analysis. In addition to fault diagnosis, it orchestrates control
Zhao, et al. Expires 31 August 2026 [Page 18]
Internet-Draft Network Management Agent Concept February 2026
plane APIs to execute self-healing operations, and integrates with
external work order systems to achieve closed-loop incident
resolution.or self-healing actions and integrates with external
work order systems to enable closed-loop incident resolution.
* *Network Optimization Agent*: Focused on performance and
efficiency, this agent translates high-level optimization goals
into technical constraints, such as load thresholds or routing
policies. Leveraging traffic prediction models, it anticipates
network congestion and proactively generates strategies for
traffic engineering (e.g., pre-diversion) and dynamic energy
saving. It operates in a closed-loop manner to autonomously
execute decisions that maintain optimal network performance.
6. Security Considerations
Since networks are critical infrastructure, misoperations can have a
significant impact on them. Therefore, NMAs shall meet the following
security and reliability requirements:
1. Support multi-factor authentication mechanism for sensitive
operations. For operations involving network configuration
changes or those that pose significant risks to network operation
security, a manual confirmation mechanism must be introduced, and
multiple authentication methods such as passwords and dynamic
tokens shall be used to ensure operation security.
2. Support circuit breaker mechanism. When abnormal results occur
during the execution of an NMA task, it shall provide error
prompts and transfer the task directly to manual control for
handling.
3. Support rollback mechanism. After the execution of an NMA task
is completed, it shall support operation rollback to restore the
network configuration.
4. Support data security and privacy protection mechanism. It shall
support the encryption of sensitive data such as network
configurations and user behavior logs; support user permission
division, and set differentiated data access permissions for
different users.
5. Support operation permission control mechanism. For different
application scenarios, the minimum permissions required to
perform tasks in the scenario shall be set. For example, a fault
handling NMA may query data such as topology resources and
performance, but shall not have permission to perform service
configuration operations.
Zhao, et al. Expires 31 August 2026 [Page 19]
Internet-Draft Network Management Agent Concept February 2026
7. IANA Considerations
This document has no requests for IANA action.
8. Appendix：Definition of L0~L5 levels in Autonomous Network
Table 1 summarizes the Autonomous Network (AN) levels defined in TM
Forum IG1230 [TMF-IG1230]. It illustrates that current IETF
automation frameworks, such as [RFC8969], primarily enable Level 3
(Partial Autonomy) by utilizing data models (YANG) to enforce pre-
defined policies.
Zhao, et al. Expires 31 August 2026 [Page 20]
Internet-Draft Network Management Agent Concept February 2026
+=====+===============+=================+==========================+
|LEVEL|NAME |DESCRIPTION (CORE| HUMAN VS. MACHINE ROLE |
| | |CHARACTERISTICS) | |
+=====+===============+=================+==========================+
|L0 |Manual |Fully manual | Human does everytding. |
| | |processes. No | |
| | |automation. | |
+-----+---------------+-----------------+--------------------------+
|L1 |Assisted |System provides | Human makes all |
| | |tools | decisions; tools assist. |
| | |(dashboards, | |
| | |alarms). | |
+-----+---------------+-----------------+--------------------------+
|L2 |System-assisted|Automation of | Human initiates tasks; |
| | |single tasks/ | system executes. |
| | |scripts witdin a | |
| | |specific domain. | |
+-----+---------------+-----------------+--------------------------+
|L3 |Partial |Closed-loop | "Human-in-the-Loop": |
| |Autonomy |automation based | Humans define rules/ |
| | |on pre-defined | models and monitor; |
| | |policies/models | system executes and |
| | |witdin a domain. | reports exceptions. |
+-----+---------------+-----------------+--------------------------+
|L4 |High Autonomy |Cross-domain/ | "Human-on-the-Loop": |
| | |cross-layer | Humans define high-level |
| | |context analysis | intents; system self- |
| | |and closed-loop | configures and heals. |
| | |optimization | Human only intervenes on |
| | |based on Intents.| system failure. |
+-----+---------------+-----------------+--------------------------+
|L5 |Full Autonomy |Self-evolving, | "Human-out-of-the-Loop": |
| | |self-optimizing, | System requires no human |
| | |fully driverless | intervention for |
| | |operations. | business goals. |
+-----+---------------+-----------------+--------------------------+
Table 1: Autonomous Network Levels (L0-L5)
Figure 5 depicts the ‘Intent-Awareness-Analysis-Decision-Execution
(IAADE)’ control loop AN architecture, highlighting the evolution
from the rule-based automation of Level 3 to the intent-driven, AI-
powered autonomy of Level 4, which is the focus of this document.
Network Management Agent can serve as an augmentation layer,
enhancing network management automation and orchestration
capabilities through natural language intent translation, cross-
vendor semantic bridging, and knowledge codification. In this
context, Agents focus on decision support and workflow orchestration,
Zhao, et al. Expires 31 August 2026 [Page 21]
Internet-Draft Network Management Agent Concept February 2026
while critical configuration changes continue to follow manual
approval and transactional execution mechanisms via existing
deterministic protocols (e.g., NETCONF), striking a balance between
automation efficiency and operational certainty.
+----------+
| INTENT |
| (Goal) |
+----------+
|
| "What to achieve"
v
+-------------+ +------------+ +------------+ +-------------+
| AWARE | | ANALYZE | | DECIDE | | EXECUTE |
| | -> | | -> | | -> | |
| (Awareness) | | (Analysis) | | (Decision) | | (Execution) |
+------+------+ +------+-----+ +------+-----+ +------+------+
| | | |
+------------------+-----------------+-----------------+
|
v
+--------------+
| NETWORK |
+--------------+
Figure 5: IAADE Control Loop for Autonomous Networks
9. References
9.1. Normative References
9.2. Informative References
[I-D.irtf-nmrg-ai-challenges]
François, J., Clemm, A., Papadimitriou, D., Fernandes, S.,
and S. Schneider, "Research Challenges in Coupling
Artificial Intelligence and Network Management", Work in
Progress, Internet-Draft, draft-irtf-nmrg-ai-challenges-
03, 4 March 2024, <https://datatracker.ietf.org/doc/html/
draft-irtf-nmrg-ai-challenges-03>.
[I-D.kdj-nmrg-ibn-usecases]
Yao, K., Chen, D., Jeong, J., Wu, Q., Yang, C., and L.
Contreras, "Use Cases and Practices for Intent-Based
Networking", Work in Progress, Internet-Draft, draft-kdj-
nmrg-ibn-usecases-01, 8 July 2024,
<https://datatracker.ietf.org/doc/html/draft-kdj-nmrg-ibn-
usecases-01>.
Zhao, et al. Expires 31 August 2026 [Page 22]
Internet-Draft Network Management Agent Concept February 2026
[LLM-powered-autonomous-agents]
Weng, L., "LLM Powered Autonomous Agents", 23 June 2023.
[RFC7575] Behringer, M., Pritikin, M., Bjarnason, S., Clemm, A.,
Carpenter, B., Jiang, S., and L. Ciavaglia, "Autonomic
Networking: Definitions and Design Goals", RFC 7575,
DOI 10.17487/RFC7575, June 2015,
<https://www.rfc-editor.org/rfc/rfc7575>.
[RFC7576] Jiang, S., Carpenter, B., and M. Behringer, "General Gap
Analysis for Autonomic Networking", RFC 7576,
DOI 10.17487/RFC7576, June 2015,
<https://www.rfc-editor.org/rfc/rfc7576>.
[RFC8969] Wu, Q., Boucadair, M., Lopez, D., Xie, C., and L. Geng, "A
Framework for Automating Service and Network Management
with YANG", RFC 8969, DOI 10.17487/RFC8969, January 2021,
<https://www.rfc-editor.org/rfc/rfc8969>.
[RFC9315] Clemm, A., Ciavaglia, L., Granville, L. Z., and J.
Tantsura, "Intent-Based Networking - Concepts and
Definitions", RFC 9315, DOI 10.17487/RFC9315, October
2022, <https://www.rfc-editor.org/rfc/rfc9315>.
[TMF-AN-journey-guide]
Tansuthepverawongse, Boonchoung., "AN Journey Guide
Autonomous Networks L4 industry blueprint-high-value
scenarios", June 2024.
[TMF-IG1230]
McDonnell, K., Machwe, A., Milham, D., O’Sullivan, J.,
Clemm, A., and J. Niemöller, "Autonomous Networks
Technical Architecture", TMF IG1230, December 2022.
Authors' Addresses
Xing Zhao
CAICT
Beijing
China
Email: zhaoxing@caict.ac.cn
Minxue Wang
China Mobile
Beijing
China
Email: wangminxue@chinamobile.com
Zhao, et al. Expires 31 August 2026 [Page 23]
Internet-Draft Network Management Agent Concept February 2026
Bo Wu
Huawei
China
Email: lana.wubo@huawei.com
Daniele Ceccarelli
Cisco
Email: dceccare@cisco.com
Haomian Zheng
Huawei
China
Email: zhenghaomian@huawei.com
Jin Zhou
ZTE
China
Email: zhou.jin6@zte.com.cn
Zhao, et al. Expires 31 August 2026 [Page 24]