CATALIST

Session: Coordinating Agent-to-Agent List of Efforts (CATALIST) BoF
Date: IETF 121, Dublin, Ireland
Chairs: Adrian Farrel, Ying-Zhen Qu
Area Director: Andy Newton

Summary

The CATALIST Birds of a Feather (BoF) session was held to coordinate the various efforts within the IETF related to AI agent-to-agent (A2A) communications. This BoF was not working-group forming and did not aim to produce a charter. Instead, it served as a forum to summarize ongoing side meetings and identify potential work items for existing or new working groups. The scope was strictly limited to A2A communication (architectures, protocols, security, discovery) and explicitly excluded "AI for Networking" and "Networking for AI."

Key Discussion Points

1. Presentation of Ongoing Efforts

3a. AI Agent Protocols
Presenter: Cullen Jennings
Cullen Jennings discussed efforts to define a charter for service invocation and session protocols (e.g., Model Context Protocol - MCP). Key areas include human-in-the-loop approvals, token delegation, and mapping work to existing IETF groups. The goal is to refine a strawman charter for a potential working-group forming BoF in Vienna.
3b. AI Networks
Presenter: Arashmid Akhavannasab
Arashmid Akhavannasab proposed a consumer-provider overlay architecture. He categorized work into agent-specific items (communication protocols, session management) and common ecosystem items (discovery, authorization). He suggested that while the Art area is a good starting point, some work should be dispatched to existing groups.
3c. AI Agent Discovery
Presenter: Nick Williams
Nick Williams highlighted the urgency of standardizing discovery to avoid fragmentation and proprietary silos. He advocated for DNS-based discovery (e.g., DNS-AID) to leverage existing federated models and operator sovereignty. He noted that while DNS is excellent for naming, it has limitations regarding metadata and complex search.
3d. AI-Agent Network Requirements and Protocols (ANP)
Presenter: Kehan Wu
Kehan Wu focused on mobile network use cases (e.g., 6G, embodied AI, autonomous vehicles). Key requirements identified include agent identity, delegated authorization, and new session semantics for non-deterministic interactions.
3e. DMSC (Dynamic Multi-agent Secured Collaboration)
Presenter: Huiyuan (Adrian) Li
The DMSC effort focuses on a manageable infrastructure using agent gateways. This approach distinguishes between server-side and client-side agents, utilizing semantic routing and natural language usage in collaboration protocols.
3f. Agent Communications (private) Internetwork Protocol
Presenter: Toerless Eckert
Toerless Eckert proposed ACIP, focusing on intermediate node processing (load balancers, firewalls) as a first-class citizen in the protocol design. The goal is to allow high-speed, hardware-based processing of agent policies and name-based forwarding.
4. Work outside the IETF
Presenter: Suresh Krishnan
Suresh Krishnan summarized external activities in the Linux Foundation (Agentic AI Foundation, MCP), IEEE, ISO (SC 42), W3C, ITU-T (SG 17), and 3GPP. He emphasized that the IETF must collaborate with these bodies to ensure IETF-produced protocols are consumable by the broader AI community.

2. General Discussion

Definitions: Linda and others expressed a need for clear definitions and categories of "agents" to ground technical discussions.
Application-Centric View: Ted Hardie suggested the IETF should focus on how these new application types (AI) affect network characteristics (resource consumption, authentication) rather than standardizing AI itself.
Marketplaces and Incentives: Dan Druta and Mallory Knodel questioned the incentives for interoperability, noting that enterprise agents are likely to operate within managed marketplaces rather than a "wild west" discovery model.
Security: Osama, Peter, and Roberta highlighted a gap in security and threat modeling. Issues such as blast radius amplification and semantic attacks (attacking via natural language content) require specialized attention.
Technical Fit: Paul Hoffman cautioned against assuming DNS is a "discovery" tool, noting its limitations for search. David Schinazi advised that the IETF is most successful with small, tightly scoped deliverables rather than broad architectural shifts.
External Coordination: Lionel Morand (3GPP) and Arnaud Taddei (ITU-T) stressed the importance of synchronization. 3GPP is currently conducting feasibility studies and would prefer to adopt IETF solutions rather than developing their own.

Decisions and Action Items

Poll on Coordination: A poll was taken on whether the IETF should have an ongoing AI Agent-to-Agent coordination activity. A sense of the room indicated support for continued coordination, though the specific format (e.g., a formal group vs. a mailing list) was not finalized.
Consensus: No technical consensus was sought on specific protocols (DMSC, ACIP, DNS-AID, etc.) as this was an informational BoF.

Next Steps

IESG Evaluation: The IESG (represented by Andy Newton) will evaluate the feedback from this session to determine if a more formal coordination structure is needed.
Dispatching Work: Proponents of specific technologies (e.g., discovery, session management) are encouraged to use existing area dispatch functions (Art, Security, Internet) to find homes for well-defined work items.
Side Meetings and Interims: Toerless Eckert and others suggested moving the heavy lifting of technical discussion to interim meetings or mailing lists to avoid "side meeting fatigue" during IETF weeks.
Mailing List: Discussions will continue on the agent-to-agent@ietf.org mailing list.