The Scope Creep Automated reviews of new AI-related working group submissions

Review · November 16, 2026

AGENTPROTO Use Cases and Requirements

draft-agentic-ai-usecases-requirements -07View on Datatracker →

The Problem

AGENTPROTO is chartered to build one thing: a dialog management protocol. It carries dialog identifiers and lifecycle state across hops, trust boundaries and modalities. The charter says it does not replace MCP or A2A.

This draft is the charter's "Use Cases and Requirements" deliverable. It describes six use cases and derives 50 requirements from them. It is the only draft on the group's IETF 127 agenda, and the session is meant to pick the minimum set of use cases for the chartered scope.

The draft does not mark a minimum set.

How It Fits In

The draft defines no mechanism. Its job is to bound the protocol work.

Its own gap analysis is narrow and matches the charter. Section 6 says A2A contexts and MCP sessions are single hop, and "no identifier spans the hops". It lists nine requirements that fill that gap: CMN-11, CMN-12, B1-6 to B1-8, B3-9, B3-10, B5-2 and B5-3.

The other 41 requirements go further than that.

Fifteen carry only the Transport tag. They ask for streaming output (A1-1, B1-4), task delegation and result messages (B1-1, B1-3), task cancellation (A1-2, B1-5), bulk data transfer (A1-8) and task priority (CMN-6). Section 4.2.1 says this pattern is already "reflected in" A2A. These describe an application-layer agent protocol, which the charter says the group is not building.

Section 5 hands another 16 to the OAuth and WIMSE working groups.

Summary

The six use cases:

  • Single agent with tools (4.1): a travel app, a travel agent, an airline booking tool in another domain.
  • Orchestrator (4.2): parallel delegation, and resuming a dialog after a dropped connection without booking the hotel room twice.
  • Authorization checkpoint (4.3): a bank account closure pauses for approval before moving the balance.
  • Multi-hop delegation (4.4): hiring, sourcing and screening agents in three domains.
  • Consensus (4.5): a coordinator and peer agents iterate on one problem.
  • Mediator (4.6): a gateway in front of tools and agents, such as a mobile operator's agent gateway.

The dialog requirements are the strongest part. B1-7 says a resuming party must be verified as the one that was interrupted, with current state, on a dialog that is not terminated or revoked. B1-8 and CMN-12 require telling "check failed" apart from "check could not be completed". Section 4.2.1 says resuming a dialog does not justify resending an operation. B3-9 allows a new linked dialog at a domain boundary.

What It Says About Authorization

Section 5 is three paragraphs. It says the authorization, delegation and identity requirements "are expected to be addressed in the OAuth and WIMSE Working Groups". It cites no OAuth or WIMSE document. The reference list has none.

So do the requirements match what those groups are building? Not all of them.

Agents issue credentials. B3-3 requires that a receiving agent can verify the credential "was issued by the delegating agent". B3-4 derives authorization from "the authorization issued by the initiating agent". In Identity Chaining and Transaction Tokens, an authorization server or token service issues each token. B3-3 describes an agent-signed attenuation chain. That is a design choice written as a requirement.

Approval travels in-band. B2-3 defines approve, deny and approve-with-changes as protocol messages. B2-5 says the response authorizes the action. The orchestrator relays the request to the user (4.3.1). AIMS Section 10.8 says the opposite: confirmation collected during a task is not authorization and must be bound to a grant from the authorization server. The draft assigns B2-2 to B2-6 to OAuth, but the text says "the protocol is required to define" the checkpoint messages. Which protocol?

Revocation propagates down the chain. CMN-10 wants a revoked authorization to reach every agent acting under it. RFC 7009 revokes a token at one authorization server.

Agent identity is missing. The charter's first key consideration is agent identity distinct from user identity, with independent revocation. The abstract promises "agent identity and authentication". No requirement says it. CMN-2 asks only for mutual authentication.

Section 5's list also skips A1-9, the delegated credential an agent presents to a tool.

Two of the authors have a companion draft, draft-reddy-wimse-oauth-dialog-binding, that binds the dialog context to WIMSE and OAuth credentials. This draft does not reference it.

Interoperability

Requirements that could be removed

  • The Transport-only requirements that restate A2A and MCP features.
  • Section 4.5. It has no requirements table and adds nothing to dialog management.
  • B3-6, a signed per-hop provenance record. The charter puts audit data structures out of scope, and Section 7 agrees.

The charter also asks for deployment models. The draft gives one sentence in Section 4.

Writing Quality

The examples are concrete and the requirement tables are readable.

There are loose ends. Section 8 points to "Identity, Authentication, and Delegation requirements". The tags are Dialog, Transport, Security and Authentication. CMN-4 and CMN-5 overlap, and A1-3 defines a third error list. Every reference is dated "n.d.".

Authors

From their Datatracker profiles:

  • Tirumaleswar Reddy.K (Nokia). 39 RFCs. 49 active drafts, mostly post-quantum work in TLS, IPsec, JOSE and COSE, plus three WIMSE individual drafts.
  • Zaheduzzaman Sarker (Nokia). 7 RFCs. 8 active drafts. Chair of TSVWG and TIPTOP.
  • Kehan Yao (China Mobile). 1 RFC. 13 active drafts. AGENTPROTO secretary and DAWN tech advisor.

All three are also authors of the group's architecture draft, draft-manyfolks-agentproto-arch.

Questions for the Working Group

  • Is the minimum set the nine dialog requirements in Section 6? If not, what else, and why?
  • Do the task messaging requirements stay, given the charter says the protocol does not replace A2A or MCP?
  • Have the OAuth and WIMSE groups been asked to take the 16 requirements in Section 5? Who tracks them?
  • Should B3-3 and B3-4 stop assuming agents issue credentials?
  • Does a checkpoint approval authorize anything by itself, or must it become a grant at an authorization server?
  • Where does the charter's agent identity requirement go?