Markdown Version | Transcript | Session Recording | Session Materials

ANIMA

Summary

The ANIMA Working Group met to discuss the status of its active draft pipeline, progress on constrained security and signaling protocols, and future directions for the working group. With core Autonomic Network Infrastructure (ANI) specifications like BRSKI nearing completion in the IESG/RFC Editor queue, the group is transitioning to focus on new work. Key themes of the meeting included rechartering to address AI-powered Autonomic Service Agents (ASAs), introducing "agentic AI" and "intent-based networking" into the ANIMA framework, and managing network-level admission/access control for distributed AI agents.

Key Discussion Points

1. Administration & Document Status

• BRSKI Cluster Status: Toerless Eckert noted that several core BRSKI documents are progressing through the RFC Editor queue. In Jabber/chat, Steffen Fries and Mahesh Jethanandani discussed how draft-ietf-anima-brski-prm (BRSKI-PRM) was pushed into the RFC Editor queue now that its dependency, rfc8366bis, is in the IESG Evaluation state.
• BRSKI Discovery: Toerless Eckert presented an update on draft-ietf-anima-brski-discovery. He incorporated feedback from Stuart Cheshire's review, clarifying the differences between the draft's goals and traditional DNS-SD (fully automated client implementation vs. human-in-the-loop, and target optimization for highly constrained devices).
• AI for Draft Editing: Toerless mentioned using Claude to proofread the XML/Markdown of the draft. Stuart Cheshire supported this strategy, noting that AI tools are exceptionally valuable for proofreading spelling, grammar, and general readability before submitting to the mailing list, allowing participants to focus purely on technical content.
• Working Group Rechartering: The chairs presented a proposal to recharter ANIMA to explicitly support AI-powered network agents (AI-ASAs) and agentic network automation. Sheng Jiang clarified that "network agents" refers strictly to agents serving networking tasks (device configuration, management, operation) rather than generic application-level AI.

2. Constrained BRSKI and Join Proxy

• Speaker: Esko Dijk

• Slides: 01 cBRSKI and Join Proxy

• Associated Drafts: draft-ietf-anima-constrained-voucher, draft-ietf-anima-constrained-join-proxy

• Updates on Constrained Voucher:

  • Updated the join proxy discovery methods using a simplified, more compact CoAP Link Format discovery.
  • Added default DTLS 1.2 and DTLS 1.3 cipher suite details to improve out-of-the-box interoperability.
  • Esko shared details on his Hackathon work using a simulated environment (running open-source components of Thread/6LoWPAN and Java-based web services) to validate the end-to-end bootstrap flow.
  • A detailed Shepherd review was received. Esko is planning to move general IoT post-onboarding device management considerations out of this document to make the draft strictly focused on onboarding.

• Updates on Constrained Join Proxy:

  • Aligned terminology with updated onboarding specs.
  • Provided stricter guidance on mode selection (stateless vs. stateful).
  • Refined the definition of the JPY protocol scheme and shortened the default multicast discovery message sizes in cooperation with the CoRE Working Group.

3. The Constrained GRASP (cGRASP)

• Speaker: BUPT Presenter

• Slides: 03 The Constrained GRASP (cGRASP)

• Associated Draft: draft-ietf-anima-constrained-grasp

• Key Updates in v03:

  • Simplified the base profile by deferring support for the standard GRASP M_WAIT message to future extensions. Since CoAP-based state management is complex over long periods, the base cGRASP profile expects nodes to rely on negotiated timeouts or terminate and restart sessions instead of waiting.
  • Redesigned the discovery relay mechanism. CoAP exchange states are now kept local to each relay hop (hub), while the cGRASP session state remains stable end-to-end.
  • Migrated the prototype implementation from TCP-based GRASP to Python/AIOCoAP (running over UDP). The prototype successfully validated multi-link relaying, unicast negotiation, synchronization, flooding, and loop-count handling.
  • The authors plan to open-source the prototype once licensing approvals are cleared.

4. Autonomic Traffic Management in Artificial Intelligence Data Centers

• Speaker: Sheng Jiang

• Slides: 04 Autonomic Traffic Management in Artificial Intelligence Data Centers

• Associated Draft Context: Formerly independent, recently submitted with an ANIMA-oriented focus.

• Discussion:

  • Sheng described the unique traffic challenges in AI Data Centers (AIDC). Highly synchronized workloads generate extreme burstiness, network hotspots, and compute-idle intervals during collective communication phases (e.g., training/inference).
  • Proposed an autonomic traffic management framework that coordinates across multiple domains: applications, compute placement, packet networks, and optical paths.
  • Outlined two operational models: a fully distributed workflow utilizing ASAs negotiating via GRASP, and a hybrid orchestration model combining global visibility (centralized controllers) with low-latency local reaction (distributed ASAs).
  • Questions to the WG: Sheng asked if traffic management ASAs are within the scope of ANIMA, whether the hybrid orchestration framework is viable, and if the draft should define specific GRASP objectives (which would shift it to the Standards Track). Toerless Eckert suggested moving these questions to the mailing list due to scheduling constraints.

5. AI for Autonomic Networking and Agentic AI for Intent-Based Networking

• Speaker: Toerless Eckert

• Slides: 05 AI for Autonomic Networking and Agentic AI for Intent-Based Networking

• Discussion:

  • Discussed the gap between "self-driving networks" (which are typically just highly automated centralized NOCs running AI) and truly autonomic, distributed networks.
  • Explored how Agentic AI can safely operate within the ANI framework. Strategies include:
    1. Letting LLMs/agents execute autonomic control loops directly, relying on the robust security/isolation of the Autonomic Control Plane (ACP) to survive agent failures or misconfigurations.
    2. Restricting LLMs to safe parameter optimization (e.g., traffic engineering).
    3. Using LLMs as development and deployment assistants to automatically synthesize and verify lightweight, deterministic ASAs.
  • Explained how rechartering would allow dynamic, lifecycle-managed ASAs to be safely downloaded and run in the network (or on provisioning PCs in POPs for legacy network environments).

6. Problem Statement and Gap Analysis for Autonomic Networking with AI-ASA

• Speaker: China Unicom Presenter

• Slides: 06 Problem Statement and Gap Analysis for Autonomic Networking with AI-ASA

• Discussion:

  • The presentation identified gaps between traditional ASAs and AI-powered ASAs (AI-ASAs) operating over the ANI.
  • Gaps Identified:
    1. Data-Intensive Communication: AI-ASAs exchange large contexts (e.g., LLM prompts, reasoning traces, real-time streams). While GRASP is suitable for discovery and negotiation, bulk data transport might require handshaking over GRASP and shifting to TCP, QUIC, or HTTP.
    2. Dynamic/Unregistered Objectives: Unlike traditional static GRASP objectives, AI-ASAs generate dynamic objectives. A common semantic model combining machine-readable structures with natural language is needed to ensure interoperability between agents from different developers.
    3. Trust & Security: Mechanisms must be introduced to address AI-specific risks, such as model errors or hallucinations, before agents deploy configurations.

7. GRASP Through Routers: A Problem Statement

• Speaker: Titouan Soulard (Grasp Routers Presenter)

• Slides: 07 GRASP through routers: a Problem Statement

• Discussion:

  • Titouan presented a scenario where GRASP is used to orchestrate a cloud environment across multiple administrative domains. In this setup, operators lack control over intermediate routers, meaning multicast GRASP discovery cannot cross subnets because the routers do not support GRASP.
  • Proposed a "relay node" objective. Relays in each subnet would discover each other (via pre-configuration) and tunnel multicast discovery messages over unicast across the non-GRASP routers.
  • In the chat, Esko Dijk asked if this model shifts GRASP from "zero-config" to "minimal-config". Titouan replied that while a one-relay-per-subnet setup requires minimal configuration (relays know each other but are discovered dynamically by clients), it is far simpler than an "omniscient" configuration where every node must know a remote relay.
  • Toerless Eckert suggested this work aligns well with proposed rechartering and a potential "ACP-free ANI" architecture.

8. Network Admission of AI Agent Instances & Scope-Down Access Control

• Speaker: Chao Zhou

• Slides:

  • 08 Use Cases and Requirements for Network Admission of AI Agent Instances
  • 09 Campus Agent Identification and Scope-Down Access Control

• Discussion:

  • Chao presented the security and operational problems that arise when multiple AI agents or processes share the same IP address on a device. Currently, the network cannot differentiate traffic between different agents or attribute traffic to specific agent instances.
  • Application-layer authorization occurs too late—after network access has already been exposed.
  • Chao proposed a "verify, bound, and enforce" model to couple cryptographic agent instance identity with network admission.
  • Additionally, because AI agents dynamically select tools and external targets based on their tasks, the network must enforce "scope-down" access control. This restricts an agent's network privileges to the intersection of its base credentials and the minimal requirements of its current, active task.

Decisions and Action Items

• Document Progress:
  • draft-ietf-anima-brski-discovery will proceed toward Working Group Last Call (WGLC) once Toerless Eckert addresses the early reviews received from other areas.
• cBRSKI Drafts:
  • Esko Dijk to submit formal responses to the Shepherd review comments on draft-ietf-anima-constrained-voucher.
  • Esko Dijk to excise post-onboarding IoT device management considerations from draft-ietf-anima-constrained-voucher and potentially move them to other working groups (e.g., ACE or LAMPS).

Next Steps

• ANIMA Rechartering:
  • Chairs to refine the proposed charter text based on meeting feedback and present it to the Area Director (AD) and the Working Group.
  • Members with existing/recent drafts (such as DNS discovery over GRASP, ACP-free ANI, and AI-ASA use cases) are urged to champion their work on the mailing list to sync with the new charter scope.
• Mailing List Discussions:
  • Debate on draft-ietf-anima-constrained-grasp v04 updates (especially discovery transaction lifecycles and cGRASP profile scope).
  • Discussion on the architectural alignment and security properties of "GRASP through routers" (07 GRASP through routers: a Problem Statement).
  • Scoping of Sheng Jiang's AI Data Center traffic management draft and Chao Zhou's agent admission and scope-down access control problem statements.

Related Documents

draft-ietf-anima-brski-discovery, draft-ietf-anima-brski-prm, draft-ietf-anima-constrained-grasp, draft-ietf-anima-constrained-join-proxy, draft-ietf-anima-constrained-voucher