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

Update review · November 10, 2026

Agent Authorization Use Cases and Gap Analysis Update

draft-chen-oauth-agent-authz-use-cases -05View on Datatracker →

Status: Individual draft (not adopted), Informational. Revision -05 published 27 October 2026, 16 days after -04. 51 pages, up from 48. 14 use cases, up from 13. Same six authors.

This revision does one thing. It adds a new Use Case 2, "Daily Small-Amount Orders," and renumbers everything after it. No other text changed.

The previous review ended by asking whether the authors intend to converge. Three revisions in a row have now added use cases without removing any.

What got fixed

Nothing. Every item in the previous review's feedback list is still open.

  • The RFC 8693 error is still there. UC8 (was UC7) still says there is "no standard for how to nest" act claims. UC13 (was UC12) still says the opposite, correctly.
  • No out-of-scope section. The OS resources and no-authorization-server use cases are retained as UC5 and UC6.
  • The MCP client measurement is still uncited. Eight clients, no names, no harness, no data.
  • RFC 9470, draft-ietf-oauth-first-party-apps, and draft-parecki-oauth-global-token-revocation are still not cited.
  • The editorial list is untouched. The uncited RFC 9486, the wrong CIMD authors in Appendix A, the abstract's dangling "these," the Section 7 comma splice, the stray "MUST," and the missing affiliations all survive.

The renumbering introduced one new defect. UC9 still says "Unlike a delegation chain (Use Case 7)." Use Case 7 is now the MCP first-connection scenario. The delegation chain is Use Case 8.

The new use case

UC2: Daily Small-Amount Orders. Dave tells his assistant to handle his daily errands. He approves one grant for the day: read, draft, and spend up to $30 per transaction and $100 total, across food delivery, travel, and household services. The grant expires at midnight. A $24 dinner order goes through. A $38 rideshare exceeds the per-transaction cap and is held for Dave's approval.

The draft calls this "amortized consent." It lists six gaps: no human-readable intent at grant time, no tiered action model, no amortized-authentication model, no grant-level revocation, no standard spend-cap semantics, and no binding of grant bounds to execution-time enforcement.

So is this new? Mostly not. UC1 already covers spending limits, intent lost at grant time, and "No Standard Binding of Pre-Authorized Constraints to Execution." UC2 restates three of UC1's gaps with a different example.

Two of the six gaps are stated as absent when the mechanism exists.

"No amortized-authentication model" is wrong. The draft says "OAuth assumes interactive, per-session approval; there is no notion of one strong authentication covering a bounded grant lifetime." That is what a refresh token with a fixed expiry is. The user authenticates once, and the client acts until the grant ends. The very next use case in the document says "Refresh Tokens are well-suited for persistent delegation." The two statements sit one page apart.

"No grant-level revocation" overlooks existing work. RFC 7009 says revoking a refresh token should also invalidate the access tokens issued from the same grant. Grant Management for OAuth 2.0 from the OpenID FAPI working group defines a grant_id and an endpoint to revoke the whole grant. Neither is mentioned. Neither is the authors' own draft-chen-oauth-agent-revocation.

What's actually worth keeping?

The cumulative cap across independent services. A $100 daily limit spanning a food delivery app, a rideshare app, and a household services app means someone has to hold a running total. No single resource server sees the other two. RAR can carry the limit, but it can't make three unrelated services agree on how much is left. This is the hard part of the use case, and the draft gives it one sentence in the last bullet. It should be the headline.

The tiered action ladder. "Read < draft < send < spend" is a concrete proposal for a consequence-ordered permission vocabulary. "The agent may draft but not send" is something users want to say and no two services express it the same way. Whether OAuth should standardize that vocabulary or leave it to each API is a fair question for the list. RAR deliberately left actions values to each API.

The gaps in Section 6 were not updated. UC2 adds six gap bullets and none of them reach the summary, so the summary is still seven items. Either the new gaps are covered by the existing seven, in which case the use case is redundant, or they aren't, in which case the summary is incomplete.

Background of the Authors

No change. The same six authors are listed. See the previous reviews for their backgrounds.

Suggested Feedback

  1. Ask on the list for a convergence plan before anything else. The document has gone from 11 to 13 to 14 use cases and from 39 to 51 pages across three revisions. None of the consolidation requested in two prior reviews has happened. A 16-day turnaround that adds a use case and addresses nothing else suggests the authors are accepting contributions faster than they are editing.

  2. Merge UC2 into UC1. They share the spending-limit scenario and three gaps. Keep UC2's two distinct points, the cross-service cumulative cap and the tiered action ladder, as additions to UC1.

  3. Fix the two incorrect gap statements in UC2. Refresh tokens already give one authentication for a bounded grant lifetime. RFC 7009 and Grant Management for OAuth 2.0 already cover grant-level revocation. Restate each gap as what those mechanisms don't do for agents.

  4. Make cross-service cap enforcement the point of UC2. Say who holds the running total: the authorization server, the agent, or a payment network. Each answer leads to a different mechanism, and the use case is only useful if it forces that choice.

  5. Update Section 6 or explain why UC2 doesn't change it.

  6. Fix the stale cross-reference in UC9. "Use Case 7" should be "Use Case 8."

  7. Everything from the previous review still applies. Fix UC8's RFC 8693 claim to match UC13. Cite the MCP client measurement. Engage with RFC 9470 and draft-ietf-oauth-first-party-apps. Re-scope UC14 around draft-parecki-oauth-global-token-revocation. Add the out-of-scope section. Do the editorial pass.