IETF’s agent protocol group puts identity and revocation in its requirements baseline

The IETF working group chartered to standardise how AI agents talk to each other has published its first document: a 29-page requirements baseline that makes mutual authentication mandatory, requires credentials to be checked at the moment they are used, and requires revocation to propagate through a delegation chain.

What happened

The Agent Communication Protocols working group (agentproto) published draft-ietf-agentproto-usecases-requirements-00 on 20 November 2026. The authors are Tirumaleswar Reddy and Z. Sarker of Nokia and K. Yao of China Mobile. It is an Informational working-group document, 29 pages, with an expiry of 24 May 2027.

The draft starts from six use cases — a single agent serving a client application, an orchestrator delegating to peer agents, a long-running delegated task with an authorization checkpoint, peer collaborative problem solving, cooperative reasoning and consensus, and tool or API mediation — and derives protocol requirements from them. Identity and authorization appear in both the common requirements and the delegated-task requirements.

The common table (CMN-1 to CMN-12) includes:

  • CMN-2: mutual authentication is required between all communicating parties.
  • CMN-4: structured errors must carry an authorization scope violation type when an agent attempts an action that exceeds or contradicts the scope delegated to it.
  • CMN-8: the protocol must let a party verify that the agent authentication credentials are valid at the time of use.
  • CMN-9: the protocol must support signalling that a presented credential has been revoked or is otherwise invalid.
  • CMN-10: the protocol must support revoking a previously granted authorization, “including propagating the revocation across a delegation chain so that affected agents cease to act under it”.
  • CMN-12: before accepting a change to a dialog, a participant must be able to verify that the change is authorized at that time, and must reject it if the check fails or cannot be completed.

The authorization-checkpoint use case adds its own requirements. A delegated agent can pause and ask the orchestrator to authorize a specific action, describing the action and its effect (B2-2). The response can approve, deny or approve with modified parameters (B2-3). A timeout must be treated as unresolved and halt the affected subtask (B2-4). An approval covers only the action in the request it answers, not a different or later action (B2-5), and authorization must be available at the granularity of a specific operation rather than a broad pre-existing grant (B2-6).

Two scoping decisions are as important as the requirements. Section 5 hands the mechanics of authorization, delegation and agent identity to the OAuth and WIMSE working groups; this draft only says the agent protocol has to carry and check the results. Section 7 puts accountability and auditing out of scope entirely, pointing to W3C PROV and Trace Context for provenance and leaving an editor’s note to add IETF audit work “once it progresses”.

The gap analysis is blunt about today’s protocols. A2A groups tasks under a context identifier and MCP groups requests into a session, but both are scoped to a single hop. When an agent delegates to another agent or calls a tool, a new context or session starts and no identifier spans the hops; cancellation is per task and per hop. That is a direct problem for anyone trying to reconstruct what an agent did across a multi-agent workflow.

Why it matters for KYA

This is the first time the IETF’s agent protocol work has produced a citable list of what an agent protocol must do about identity. That matters because “KYA” has mostly been a policy demand — verify the agent, bind it to a principal, scope it, revoke it — without a requirements document that a protocol designer or a buyer can point at. This draft gives each demand a requirement ID.

CMN-10 is the one to read twice. Most agent identity products stop at issuance: they mint an identity, put it in a directory and call it done. The draft says revocation has to travel down the delegation chain so that downstream agents stop acting. That is the difference between revoking a credential and revoking authority, and it is where most real deployments are weakest. The same logic runs through CMN-8, which asks for validity at the time of use rather than at the time of issuance.

The checkpoint requirements are a useful corrective to coarse consent. B2-5 says an approval authorizes only the action it answers; B2-6 says a sensitive action should be authorized individually. B2-4 makes a missing answer fail closed. Those three together rule out the pattern where a user approves “the agent” once and every later action inherits the approval.

The division of labour is a practical guide to what to watch. agentproto is a dialog and transport protocol; it will not define the credential format, the mandate or the delegation token. Those are expected from OAuth and WIMSE. If you are building a KYA control plane, the artifact you verify will come from those groups; agentproto is where the plumbing that carries and checks it will be specified. And because accountability and auditing are out of scope, an audit trail remains your engineering problem, not something this protocol hands you.

The limits are real. This is a first working-group revision, Informational, with no implementations cited. “Required” here means a requirement in a proposal, not a rule anyone has to follow. The text can change substantially before it becomes an RFC, and some of the harder questions — how a verifier actually obtains revocation status across organisational boundaries, for instance — are deferred rather than answered.

For practitioners, the practical use is as a checklist against your own agent stack: do your agents mutually authenticate, or does only the caller authenticate? Can you distinguish an authorization failure from an authentication failure in your logs? Does your revocation story reach sub-agents and tools, or stop at the first hop? Can you authorize a single sensitive action without re-approving the whole task? If the answer to any of those is no, the gap is in your deployment, not in the draft.

What to watch

  • Whether CMN-10 survives review intact. Cross-chain revocation is expensive to implement and easy to water down into per-hop notification.
  • Whether the OAuth and WIMSE work produces a delegation artifact this draft can reference by name before the document expires in May 2027.
  • Whether A2A and MCP adopt a cross-hop dialog identifier, or leave correlation to each deployment.
  • Whether an accountability section appears in a later revision. An editor’s note is not a commitment.

Sources

  • IETF Datatracker, draft-ietf-agentproto-usecases-requirements-00, “Agentic AI Use Cases and Requirements”, 20 November 2026 — datatracker.ietf.org
  • IETF, draft text (HTML) — ietf.org
  • IETF Datatracker, Agent Communication Protocols (agentproto) working group — datatracker.ietf.org
  • IETF 127 proceedings, San Francisco, 14–20 November 2026 — datatracker.ietf.org

Related on kya.fi: IETF approves agentproto charter, delegation chain, revocation, audit trails for agents, and the standards and regulation tracker.


Know Your Agent (KYA) explains agent identity, verification and accountability. This is an explainer, not legal or compliance advice — see our Sources & methodology. New to KYA? Start with What is Know Your Agent? and the glossary.