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

Update review · October 26, 2026

Agent Authorization Use Cases and Gap Analysis Update

draft-chen-oauth-agent-authz-use-cases -04View on Datatracker →

Status: Individual draft (not adopted), Informational. Revision -04 published 11 October 2026. 48 pages, up from 40. Six authors, up from five. 13 use cases, up from 11. Seven gaps in the summary, up from six.

The previous review asked for one thing above all: draw boundaries and compress. This revision did the opposite. It added two use cases, one gap, one author, and eight pages. The two additions are the most carefully written material in the document, which makes the trajectory problem harder to argue, not easier.

What got fixed

Of the ten items in the previous review's feedback list, the editorial ones got partial attention and the substantive ones mostly did not.

The CIMD reference name is correct now. Section 5.3 cites draft-ietf-oauth-client-id-metadata-document instead of the nonexistent draft-ietf-oauth-client-metadata. The conflation with RFC 7591 in the body of that section is unchanged. The new UC6, though, cites CIMD and Dynamic Client Registration as two separate mechanisms, so the authors clearly know the difference.

RFC 2119 and RFC 8174 are gone from the references. The stray uppercase "MUST" in UC10's gap analysis ("it MUST require one access token at a time") is still there, so the keyword now has no definition at all.

Transaction tokens are cited as a specification, in one place. The new UC12 cites draft-ietf-oauth-transaction-tokens and the Fletcher chaining profile directly. Section 5.4 still describes transaction tokens as "a common architectural pattern, often implemented using JWTs." Same document, two different levels of engagement with the same draft.

The nested bullet markup was normalized. Every "What's Missing" list is now consistently indented one level under its heading. This is real cleanup across all 13 use cases. The one malformed marker called out last time, **What's Missing (The Gap) in the DNSSEC use case, survived the pass.

Small items: the missing space in "morganLR,Eric Leleu" is fixed. The DS automation draft is now a proper reference rather than an inline filename. RFC 9449 and RFC 8628 moved from inline citations into the reference list.

What did not get fixed

  • No out-of-scope section. UC4 (OS resources) and UC5 (no authorization server) are retained unchanged.
  • The abstract still opens with "these emerging agent-based use cases." One line, still dangling.
  • The comma splice in the Durable Record bullet of Section 7 is unchanged.
  • The RFC 8693 nested act claim error in UC7 is unchanged. "There is no standard for how to nest these claims" is still there. More on this below, because the new UC12 contradicts it.
  • draft-parecki-oauth-global-token-revocation is still not cited. UC13 (security incident response) is word-for-word the same as UC11 was.
  • RFC 9470 and draft-ietf-oauth-first-party-apps are still not cited. This matters more now, because the new UC6 is entirely about resource server challenges.
  • Appendix A still duplicates Section 10, and it grew. It now has its own entry for CIMD that lists the authors as "Brossard, D. and T. Lodderstedt" dated October 2024. Section 10 lists the same draft correctly as Parecki and Smith, July 2026. The document cites one draft two ways with two different author lists.
  • Yuning Jiang still has no affiliation, and the new sixth author has none either.
  • Requirements are still aspirations. No testability pass was done.

One new defect: RFC 9486, "IPv6 Options for In Situ OAM," was added to the reference list and is cited nowhere in the body. This is almost certainly a copy error from another draft.

The two new use cases

UC6: First Connection with a Front Channel the Client Does Not Use. This is the MCP authorization flow. A client hits a resource server for the first time, gets a 401 with a WWW-Authenticate pointing at RFC 9728 protected resource metadata, and the authorization server supports the code flow with PKCE. The user is sitting right there. The question is whether the client starts the flow, and if it doesn't, whether the resource server ever learns why.

The draft frames three gaps. "No expectation to act" says the specs describe what a client can do after a challenge but never when it should. "No disposition" says a client that gives up looks identical at the resource server to one that is mid-sign-in. "No presence signal" says there's no way to tell the resource server a user is available.

It backs this with a measurement: eight shipped MCP clients were pointed at a resource server that returned a 401 with resource_metadata. In one-shot mode, one ran the grant on its own, one did so only with a per-server config flag, and the rest stopped after discovery or did nothing. All eight could complete the grant through their own sign-in steps. This is the first empirical data in the document, and it is uncited. Which clients, what harness, where is the data? A use-cases draft that makes a factual claim about shipping software needs to point at something reproducible.

So is this an OAuth gap? Two of the three are not. Whether a terminal agent pops a browser without asking is a product decision, and a reasonable one to decline. OAuth has never told clients when to initiate authorization, and it would be strange to start. The "visible destination" requirement, that the client show the user which authorization server it is about to send them to, is good advice and belongs in a security considerations section of the MCP authorization spec, not in a new OAuth mechanism.

The third one, terminal disposition, is a genuinely new idea and the draft promotes it to gap number seven in Section 6. The proposal is a small enumerated set the client sends to the resource server when it won't complete the challenge: user interaction required, pending, flow not supported, refused by policy, failed. The problem is that the resource server has no channel to receive this. A client that isn't going to authorize has no reason to send another request, and a resource server that never sees a token has nothing to correlate the signal against. The draft says this would help "incident triage and audit evidence," but the audit evidence of a client that never authorized is already clear: no token was ever issued. The WG should push on who consumes this signal and what they do differently with it.

UC12: Industrial Agent for Equipment Maintenance. A plant manager asks an agent to investigate a production issue. The agent delegates diagnosis to an equipment supplier's agent under read-only authority. The diagnosis proposes an operational change, which exceeds the original grant, so the coordinating agent goes back to the manager for a second approval scoped to that one action. The approved action is delegated to an execution agent, and the supplier's diagnostic agent does not inherit it.

This is the best-written use case in the document. The gap analysis cites RFC 7662, identity chaining, transaction tokens, the chaining profile, and the authors' own collaboration draft, and then says what each one does not cover. The three gaps are framed as "incomplete" rather than "absent." It even concedes that "deployments that accept the issuing authorization server's assurance can already rely on existing trust mechanisms and policy." That is the first sentence in any revision of this document admitting an existing mechanism is sufficient for some deployments. It is exactly the register the previous review asked for.

It is also a fifth use case producing the same gap list as UC7, UC9, UC10, and UC11: attenuated multi-hop delegation, a shared task identifier, task-scoped revocation across domains. The previous review asked for those four to be compressed into one. Instead there are five. The one genuinely new requirement in UC12 is task-bound additional authorization: a mid-task approval that attaches to the original task context and extends authority for specific agents only. That is a real thing, closely related to incremental authorization, and it deserves its own entry in Table 2 rather than being spread across the existing rows.

The contradiction between UC7 and UC12

UC7 still says there is "no standard for how to nest" act claims. UC12, 15 pages later, says "Token Exchange [RFC8693] supports delegation and nested act claims identifying current and prior actors" and then makes the correct, sharper point: "Actor history in [RFC8693] does not itself prove the restrictions granted at each delegation step."

UC12 is right. UC7 is wrong. Whoever wrote UC12 read RFC 8693 Section 4.1 and the earlier author did not. Fixing UC7 to match UC12 would resolve the previous review's second feedback item in one paragraph.

Background of the New Author

Xiang Liu is listed sixth with no affiliation and a Gmail address. The Datatracker profile lookup could not be completed from this session, so RFC count and draft history are unverified. No other draft in the OAuth, WIMSE, or SCIM corpus lists this address. Given the gmail address and the absence of any other local footprint, this is most likely a first-time IETF author. The timing suggests UC6 and its MCP client measurements may be this author's contribution, which would be worth confirming on the list since the measurement methodology is the thing that needs a citation.

The five original authors are unchanged. See the previous review for their backgrounds.

Suggested Feedback

Ordered by what would most improve the document:

  1. Pick one of UC7 or UC12 as the enterprise delegation exemplar and cut the rest. UC12 is better written and has the one new requirement, task-bound additional authorization. UC7 has the Data Subject observation. Keep whichever the authors prefer, merge the other's unique point into it, and reduce UC9, UC10, and UC11 to a short note that the same requirements recur. This has now been asked for twice.

  2. Fix UC7's RFC 8693 claim to match UC12's. Nested act claims exist. The gap is that they are informational only and prove nothing about restrictions. UC12 already says this correctly.

  3. Separate the three UC6 gaps and decide which is OAuth's. "No expectation to act" and "visible destination" are client behavior guidance and belong in the MCP authorization spec's security considerations. Terminal disposition needs a stated consumer: who receives it, over what channel, and what do they do with it that they couldn't do with "no token ever arrived."

  4. Cite the MCP client measurement. Name the clients, the harness, and where the data lives. An uncited empirical claim in an Informational draft will be the first thing challenged.

  5. Engage with RFC 9470 and draft-ietf-oauth-first-party-apps in UC6. The use case is about resource server challenges and non-redirect user interaction. Both documents are directly on point and still uncited anywhere.

  6. Re-scope UC13 around draft-parecki-oauth-global-token-revocation. Unchanged from the previous review. UC13 was not touched in this revision.

  7. Add the out-of-scope section. Also unchanged. UC4 and UC5 are still presented as gaps the WG should fill.

  8. Editorial pass: remove the uncited RFC 9486; fix the Appendix A CIMD entry or, better, delete Appendix A; fix the abstract's dangling "these"; repair the Section 7 comma splice; fix the DNSSEC use case's **What's Missing marker; lowercase or remove the stray "MUST" in UC10; update Section 5.4 to cite the transaction tokens draft the way UC12 does; fill in affiliations for Yuning Jiang and Xiang Liu; normalize the acknowledgements, which now include two more first-name-only entries.

  9. On the list, ask whether the authors intend to converge. Two revisions in a row have each added use cases without removing any. If the plan is to keep adding scenarios as they come in, the WG should know that before anyone proposes adoption, because a document that grows eight pages per revision is a document that is never done.