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

Review · November 22, 2026

Framework, Use Cases and Requirements for AI Agent Protocols

draft-rosenberg-agentproto-usecases -00View on Datatracker →

The Problem

AI agents talk to users, call APIs, and hand work to other agents, often across company boundaries. This draft maps that territory into three protocol classes and lists what each one needs. It defines no protocol. Its stated goal is to "set the stage for possible standards activity at the IETF."

The key claim is that MCP, A2A and Agntcy cover only part of the requirements. The gaps it names are mostly about identity, authorization and user confirmation.

How It Fits In

This is a use cases and requirements document, which is one of the three deliverables in the agentproto charter. The working group has already adopted a document for that slot, draft-ietf-agentproto-usecases-requirements. This draft is now input to that document, not a competitor for adoption.

It is also not new text. It replaces draft-rosenberg-aiproto-framework-00 from October 2025, and the body is nearly identical. The only content change is a reworded opening paragraph in Section 2.2 that calls APIs "tools". Everything else is the pre-charter text under a new name.

That matters because the charter that was eventually approved is much narrower than this draft. The charter covers dialog identifiers, lifecycle state, and propagation of dialog context. It puts hallucination prevention, UI design and agent decision-making out of scope. Large parts of this draft land outside those lines.

It depends on no OAuth RFCs. The only references are transport and SIP-era RFCs used for an analogy. OAuth is mentioned by name throughout, but nothing is cited.

What the Draft Describes

There are three protocol classes.

User to agent. Voice, chat, SMS, video. The draft says existing protocols cover this and no new work is needed. The one open issue is authentication over channels like phone and SMS, where the user proves identity by reciting an account number or date of birth and no token is ever issued.

Agent to API. The draft asks whether OpenAPI is already enough, and answers no. It lists four gaps:

  • An agent calling APIs with a broad service account is dangerous when prompt injection can redirect it. The agent should hold a token scoped to the one user it is serving.
  • When the user authenticated by proof of knowledge, some identity service has to mint that token, and the token should record how strong the authentication was. APIs should be able to reject it and demand more.
  • Scopes need natural language descriptions the agent can read across domains, so it can explain to the user what a permission grants before asking for it. The API description the agent receives should be filtered to the granted scopes.
  • APIs with side effects need to be marked as requiring user confirmation, with a mapping from user-visible field names to the JSON payload. The confirmation is rendered by code, not by the LLM.

Agent to agent. This is where the authors' SIP background shows. The draft borrows five contact center topologies for bringing in a second agent: blind transfer, supervised transfer, sidebar, conference and passthrough. It then lists requirements for capability negotiation (does the other agent support voice, images, video), lifecycle (synchronous or asynchronous, transactional or conversational, callbacks, polling, enumerating pending tasks), context transfer, and tagging messages as meant for the user or for the invoking agent.

For identity, the invoked agent either accepts a conveyed identity from the invoking agent or re-authenticates the user itself. The draft notes that re-authentication by PII through a relaying agent leaks that PII to the relay, and asks for end-to-end encryption or an out-of-band path.

It also asks for cryptographic proof that a user confirmation really came from the user, and not from a hallucinating intermediate agent.

So What Does This Mean for OAuth?

Most of the agent-to-API requirements already have OAuth answers, and the draft doesn't acknowledge them.

  • A lower-privilege token for the user being served is Token Exchange, plus the cross-domain work in Identity and Authorization Chaining.
  • An API rejecting a token for insufficient authentication strength and telling the client what is needed is the Step Up Authentication Challenge, RFC 9470. The acr and amr claims already carry how the user was authenticated.
  • Confirming specific transaction details with the user outside the LLM is close to Rich Authorization Requests.

Two requirements are real gaps.

The first is machine-readable and human-readable scope descriptions published by the resource domain. OAuth has never standardized what a scope means. An agent that has to explain "change reservations also lets me delete them" has nothing to read today.

The second is minting a token from identity facts collected in a conversation. Rosenberg has since written this up as a concrete proposal in draft-rosenberg-oauth-aauth. That draft is the place to debate it, not this one.

The hardest question in the draft gets one sentence. How does a user grant permission to an agent whose tasks aren't an enumerated set of APIs? The draft says this "would need to be sorted out."

Interoperability and Necessity

There is nothing to implement here, so the usual interoperability test doesn't apply. The useful test is whether the requirements can drive protocol design. They can't yet.

The requirements are not stated as requirements. There are no numbered or labeled requirements anywhere. They are buried in prose as "needs to", "would ideally" and "perhaps". A working group can't check a protocol against this document.

The survey is stale. The draft says MCP is "targeted for operation within a single host and not over the Internet" and so doesn't address inter-domain security. MCP has had a remote HTTP transport and an OAuth-based authorization spec since 2025. The A2A link points at a retired Google GitHub page. The Agntcy "ACP" naming has also moved on. The conclusion that existing protocols leave "ample room" rests on this outdated survey.

Much of it is outside the charter. Sections that could be dropped or moved:

  • Section 3.2.3 and 3.3.6 on user confirmation. These are about containing hallucination, which the charter excludes. The cryptographic confirmation idea is interesting and belongs in an OAuth or WIMSE discussion.
  • Section 3.2.4 on serving different API descriptions per LLM model. The draft itself says it's not clear this is needed.
  • Section 3.2.1 and 3.3.1 on discovery. Both conclude that discovery mostly doesn't matter because a prior trust relationship is required.
  • Section 3.1 on user-to-agent channels, which the draft says needs no new work.

What is worth keeping. The five transfer topologies and the lifecycle section map directly onto the charter's dialog management deliverable. Supervised transfer, passthrough and async callback are concrete cases a dialog identifier and lifecycle model have to survive. That material should be pulled into the adopted use cases document.

Writing Quality

The prose is clear and readable. The examples are concrete: a bank, a travel site, United and Delta, an HR approval. That part is good.

The document is unfinished. There are nine "Section XX" and "Section YY" placeholders and a literal "(ref)" where the MCP citation should be. MCP, A2A, Agntcy and OAuth have no entries in the references. There is no Security Considerations section and no terminology section.

The actors are inconsistent. The agent is "AI Agent A" in the figure and "AI agent 1" in the text. The supervisor is User S in one paragraph and User B in the next, while User B means something else in the figure. The "Transactional vs. Conversational" definition describes the conversational case using the word "transactional".

There are many typos: "reduced load rates", "To creates", "prior generates", "real a one-time code", "Anthropic" listed as a model name.

Length is fine at 26 pages. A fair amount could be cut, as listed above, and the saved space spent on stating requirements explicitly.

Authors

Jonathan Rosenberg (Five9) has 72 RFCs from 1999 to 2020, including SIP (RFC 3261), ICE, STUN and TURN. Rosenberg is one of the most published authors in the IETF. Nine active drafts, split between vCon and agent work, including the agentproto architecture draft, ASIP, and AAuth.

Cullen Fluffy Jennings (Cisco) has 29 RFCs from 2002 to 2026 and is a reviewer on the ART Area Review Team. Jennings has 30 active drafts right now, mostly MoQ, plus a cluster of agentproto, OAuth and WIMSE drafts with Suhas Nandakumar.

Both are highly experienced. The rough state of this text reflects a quick resubmission under a new name, not unfamiliarity with the process.

Bottom Line

This is a positioning document from before the working group existed, republished with almost no changes. Its transfer topologies and lifecycle cases are useful input for the adopted use cases draft. Its authorization requirements mostly restate problems OAuth has already solved, with scope descriptions and conversational authentication as the two real exceptions. Read Sections 2.3 and 3.3.3, and skip the protocol survey.