Agentic AI Architectural Principles for Autonomous Computer Networks

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶

Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶

This Internet-Draft will expire on 6 May 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.¶

Table of Contents

1. Introduction

Computer networks have long pursued higher automation for self-configuration, self-healing, self-optimisation, and self-protection. Recent advances in AI, including agentic AI systems, enable distributed, goal-driven closed loops that can plan actions, invoke tools, adapt to feedback, and coordinate with peers to improve network operation.¶

Communication services may involve multiple requirements, including throughput, latency, reliability, service continuity, and efficient resource use. Fulfilling these requirements can depend on decisions made at several protocol layers and across multiple network nodes. An AI agent operating within one layer or administrative domain may have only partial visibility of the service and limited authority over the resources contributing to it.¶

AI agents can support these services by interpreting objectives, deriving requirements relevant to their operational scope, and cooperating with other agents to fulfil the service requirements.¶

Research on cross-layer agentic networking illustrates cooperation among specialised AI agents [SANET], while work on intent-aware transport intelligence for immersive applications illustrates adaptation at a particular protocol layer [XR-TRANSPORT].¶

This document proposes architectural principles for agentic augmentation of the Internet protocol suite. AI agents act through available protocol functions, management capabilities, and other tools. Their operation is bounded by policies and authorisation, preserving the interoperability and operational constraints of the underlying network.¶

2. Terminology and Conventions

AI Agent: A software entity capable of pursuing goals through capabilities such as reasoning, planning, and invoking tools to observe its operating environment and perform actions. An AI agent operates within an assigned operational scope and autonomy layer. Throughout this draft, the terms agent and AI agent are used interchangeably.¶

Agentic AI System: A system composed of one or more AI agents that performs goal-directed operations over time using available tools and feedback. Participating agents may cooperate through decomposition, delegation, negotiation, and assessment of outcomes.¶

Intent: A statement of desired objectives and associated constraints, without prescribing how to achieve them [RFC9315]. Its representation may vary according to the receiving AI agent’s capabilities and operational scope, including natural-language expressions or structured representations.¶

Intent Decomposition: The derivation of more specific objectives and constraints from a received intent to support its fulfilment.¶

Operational Scope: The particular functions, resources, and service context for which an AI agent is responsible within one or more protocol layers.¶

IP Suite Stack Interfaces: The conventional interfaces through which protocol layers provide services, together with control and management interfaces used to configure or observe network behaviour.¶

Tool: A callable capability used by an AI agent to obtain information or execute actions. Examples include telemetry retrieval, configuration APIs, diagnostic functions, knowledge bases, and network digital twins.¶

3. Baseline: IP Suite Layered Protocol Stack

Layering provides a conceptual basis for protocol interoperability. Upper layers use services provided by lower layers through defined interfaces. Deployed systems commonly represent the Internet protocol suite using Application, Transport, Internet, and Link layers. Physical-layer functions support the underlying transmission technologies. This document uses Network Access to refer to the link layer of the Internet protocol suite [RFC1122].¶

Figure 1 presents the layered baseline used in this document.¶

+----------------------+
|     Application      |
+----------------------+
|      Transport       |
+----------------------+
|       Internet       |
+----------------------+
|    Network Access    |
+----------------------+

Figure 1: IP Suite Layered Baseline¶

This layered abstraction provides a useful basis for discussing where agentic capabilities can observe network conditions and influence operation. It does not imply that every compute node implements every layer or that every layer requires an AI agent.¶

4. Proposed Agentic Augmentation of the Layered Approach

The proposed approach introduces agentic functions alongside selected protocol layers. Existing protocol mechanisms continue to provide packet transport and interoperability, while agents use available tools to observe conditions, reason about objectives, and perform permitted actions.¶

Coordination, decomposition, and delegation may be performed by participating agents. These responsibilities may be distributed among agents or undertaken by particular agents according to deployment needs. A separate agent controller is not required by the architectural model.¶

The model supports structured autonomy: agents can act independently within their associated protocol layers, operational scopes, and authorised tool access, while cooperating when an objective requires capabilities beyond their individual scope.¶

+------------------+       +----------------------------+
|   Application    |<----->|     Application Agent      |
+------------------+       +----------------------------+
                                 |              ^
                                 v              |
+------------------+       +----------------------------+
|    Transport     |<----->|      Transport Agent       |
+------------------+       +----------------------------+
                                 |              ^
                                 v              |
+------------------+       +----------------------------+
|     Internet     |<----->|      Internet Agent        |
+------------------+       +----------------------------+
                                 |              ^
                                 v              |
+------------------+       +----------------------------+
|  Network Access  |<----->|   Network Access Agent     |
+------------------+       +----------------------------+

    Horizontal: observation and permitted action through tools,
                e.g., configuration
    Downward:   intent decomposition
    Upward:     capabilities, constraints, proposals and outcomes

            Communication service intent
                         |
                         v
           +---------------------------+
           |     Application Agent     |
           +---------------------------+
                |                 ^
Communication   |                 | Feasibility,
objectives and  |                 | constraints and
constraints     v                 | outcomes
           +---------------------------+
           |      Transport Agent      |
           +---------------------------+
                |                 ^
Delivery and    |                 | Proposed network
network-service |                 | contribution and
requirements    v                 | limitations
           +---------------------------+
           |      Internet Agent       |
           +---------------------------+
                |                 ^
Requirements    |                 | Available capacity,
for relevant    |                 | link conditions and
links           v                 | outcomes
           +---------------------------+
           |   Network Access Agent    |
           +---------------------------+

5. AI Agent Functional Capabilities

An AI agent may provide the following capabilities:¶

• Intent Interpretation and Context Management: Interpret objectives and constraints and maintain relevant observations, policies, and operating context.¶

• Reasoning and Planning: Evaluate alternatives and determine candidate actions or contributions.¶

• Intent Decomposition and Delegation: Derive objectives and involve other agents when their capabilities are required.¶

• Coordination and Conflict Handling: Exchange proposals, identify incompatible objectives or actions, and negotiate or escalate unresolved conflicts.¶

• Execution and Tool Use: Invoke permitted capabilities to observe state or perform actions.¶

• Synthesis and Outcome Assessment: Combine contributions and evaluate their relationship to the received objective.¶

• Feedback and Adaptation: Monitor effects and revise plans in response to changing conditions or unsuccessful actions.¶

These capabilities are illustrative and need not correspond to separate software modules. An agent may support a subset according to its role, and feedback-based adaptation does not necessarily require online learning.¶

The AI agent’s capabilities apply to the protocol layer or layers in which it operates. Its operational scope identifies the particular functions and resources for which it is responsible, while authorised tool access determines the operations it may perform. For example, a Transport Agent may adapt transport parameters for specified connections, while an Internet Agent may manage routing functions within a specified network domain. Figure 3 provides an illustrative functional view.¶

               Intent, constraints
                       |
                       v
+--------------------------------------------------+
|                    AI Agent                      |
|                                                  |
|  +--------------------------------------------+  |
|  | Intent Interpretation and Context          |  |
|  +--------------------------------------------+  |
|                                                  |
|  +-------------------+  +---------------------+  |
|  | Reasoning and     |  | Decomposition and   |  |
|  | Planning          |  | Delegation          |  |
|  +-------------------+  +---------------------+  |
|                                                  |
|  +-------------------+  +---------------------+  |
|  | Coordination and  |  | Synthesis and       |  |
|  | Conflict Handling |  | Outcome Assessment  |  |
|  +-------------------+  +---------------------+  |
|                                                  |
|  +-------------------+  +---------------------+  |
|  | Execution and     |  | Feedback and        |  |
|  | Tool Use          |  | Adaptation          |  |
|  +-------------------+  +---------------------+  |
|                                                  |
|  Operational scope, policy and autonomy boundary |
|             apply across capabilities            |
+--------------------------------------------------+
          ^                         ^
          |                         |
          v                         v
   Available tools            Other AI agents

   Outputs include proposals, delegated objectives,
            actions and assessed outcomes

Figure 3: Illustrative AI Agent Capabilities¶

6. Challenges Motivating Interoperable Work

The following challenges motivate further research and discussion:¶

• AI Agent identification and reachability: Agents associated with different layers, platforms, and domains need to identify and reach appropriate cooperating agents.¶

• Capability discovery and availability: Selecting an AI agent requires understanding its relevant capabilities, operational scope, supported intent representations, and current availability.¶

• Partial fulfilment and failure signalling: Agents need to communicate limitations, unsuccessful actions, and changes to previously proposed contributions so that dependent plans can be reconsidered.¶

• Authorisation and delegated responsibility: Cooperation across domains requires scoped authority and clear responsibility for actions, including the handling of delegated objectives.¶

These challenges become more complex where agents operate across different trust boundaries and use heterogeneous tools.¶

7. Security, Safety, Determinism, and Accountability Considerations

Agentic AI systems introduce risks associated with tool invocation, distributed decision-making, and the interpretation of external information.¶

Security: Tool invocation chains can create opportunities for privilege escalation. Inter-agent exchanges can propagate misleading information or compromise, while persistent context and model inputs may be subject to manipulation. Authentication, scoped authorisation, and validation of external inputs are therefore relevant throughout agent operation.¶

Permitted Operations: An AI agent’s actions remain subject to its associated protocol layers, operational scope, authorised tool access, and applicable policies. Receiving a request or delegated objective does not extend the agent’s authority to access resources or perform operations.¶

Safety and Protocol Constraints: Agentic augmentation is intended to preserve protocol correctness and operational stability. Relevant safeguards include constrained action spaces, validation before execution, bounded retries, rate limits on reconfiguration, and escalation where objectives cannot be fulfilled safely.¶

Recovery mechanisms may include rollback or compensating actions where supported. Their applicability depends on the operation, since some actions cannot be fully reversed.¶

Conflict Handling: Agents may encounter competing objectives or incompatible proposed actions. Conflict handling may require negotiation, policy-based prioritisation, or escalation to an authorised decision-maker. An agent should not assume authority over resources outside its scope to resolve a conflict.¶

Accountability: Decisions and actions should be attributable to the relevant objectives, observations, policies, and authorisation context. Records of delegation and outcomes can support troubleshooting and assessment of service fulfilment without requiring a particular identifier or logging format in this document.¶

Sustainability: Inference, telemetry collection, and AI agent coordination consume compute, network, and energy resources. The benefits of adaptation should be assessed alongside these costs. Coordination frequency and decision timescales should be appropriate to the managed functions.¶

These considerations establish architectural objectives. Their effectiveness requires evaluation under realistic operating conditions.¶

8. Recommendations and Next Steps for Research Discussion

The following discussion items are proposed:¶

1. Refine terminology for operational scope, autonomy layers, decomposition, and delegation.¶

2. Develop representative examples of cooperation across layers, compute nodes, and administrative domains.¶

3. Investigate how decomposed objectives and AI agent contributions can be assessed against an overall service objective.¶

4. Identify the information and techniques needed for capability discovery, cooperation, and failure handling.¶

5. Evaluate approaches to scoped authorisation, accountability, and escalation during delegation.¶

9. IANA Considerations

This document requests no IANA actions.¶

10. Acknowledgements

TBD.¶

11. Informative References

[RFC1122]

Braden, R., Ed., "Requirements for Internet Hosts - Communication Layers", STD 3, RFC 1122, DOI 10.17487/RFC1122, October 1989, <https://www.rfc-editor.org/rfc/rfc1122>.

[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>.

[SANET]

Xiao, Y. and et. al., "SANet: A Semantic-aware Agentic AI Networking Framework for Cross-layer Optimization in 6G", IEEE Transactions on Mobile Computing, DOI 10.1109/TMC.2026.3691804, 2026, <https://doi.org/10.1109/TMC.2026.3691804>.

[XR-TRANSPORT]

Huynh, V. S. H., Qian, P., Wang, N., Udora, C., and R. Tafazolli, "User-Intent Aware Transport-Layer Intelligence for Frame Synchronisation in Multi-Party XR Application", 2024 IEEE Gaming, Entertainment, and Media Conference, DOI 10.1109/GEM61861.2024.10585795, 2024, <https://doi.org/10.1109/GEM61861.2024.10585795>.

Authors' Addresses