Review · August 26, 2026
Agent Authorization Use Cases and Gap Analysis
Status: Individual draft (not adopted), Informational, 39 pages, 5 authors. Revision -03 published 25 August 2026.
Problem Statement
This is not a protocol document. It proposes no new mechanism and has no IANA actions. It is a survey: 11 agent use cases in three categories, each with a requirements list and a gap analysis against OAuth 2.0 and its extensions, offered as a foundation for future OAuth WG work.
The genuinely novel contribution — and the thing worth engaging with — is a conceptual split the draft introduces in §2 and then uses as the spine of the whole document:
- Grant Layer Authority — what an agent is allowed to do in principle (scopes, RAR objects, established at authorization time).
- Execution Layer Evidence — a non-repudiable record, generated at the moment a specific action is taken, proving a principal consented to those concrete parameters ("the user approved booking Sunnyvale Park for $25 on this date").
The draft's thesis is that OAuth is entirely a grant-layer framework, that agents make the gap between the two layers operationally dangerous, and that the WG needs execution-layer primitives. Everything else in the document is scaffolding for that claim.
Three headline gaps are called out in §3 with ASCII diagrams:
- Authorization context gap — the user states intent at T1; the agent requests a token at T2; the consent prompt at T2 is generic and the AS has no standard way to carry or validate the original intent.
- Delegation chain gap — with A→B→C, the RS can validate only C's token, not that C acts for B acting for A under the user's authority.
- Mass revocation gap — RFC 7009 revokes one token at a time, leaving a vulnerability window where no atomic "kill switch" exists.
How the Spec Fits In
It sits upstream of the WG's chartered work rather than inside it. The OAuth charter's "Complex Delegation" item explicitly covers "authorization of automated agents working on behalf of users, including scenarios where automated agents act across multiple administrative domains." This draft is an attempt to define the requirements space for that item. That makes it in scope for the WG to consider, but as a framing document, not a deliverable that anything implements.
Dependency load is light by construction. All references are informative (correctly, for an Informational document): RFC 6749, 7009, 7523, 8693, 9396, with RFC 9449 and 8628 cited inline. It also cites six Internet-Drafts.
One new role is proposed, and it deserves attention. §2 defines Data Subject — "a natural person whom accessed data describes, and who is neither the Resource Owner nor a party to the authorization exchange." This is the only place the draft adds to OAuth's cast of characters, and it is a real structural observation: in the insurance-claim use case, the person whose medical records are read is not the resource owner and never appears in the OAuth exchange. Whether OAuth should represent that party at all is a substantive question the WG has not settled, and the draft is right to surface it. It is also the item most likely to get lost in a 39-page document.
Notable relationship to the authors' other work. Of the six I-Ds cited, three have one of this draft's authors as a co-author (song-oauth-ai-agent-collaborate-authz, ni-oauth-batch-authorization-delegation, and the transaction-token/chaining cluster is adjacent). The author team also has at least four companion proposal drafts in flight that map onto gaps identified here: draft-chen-oauth-agent-revocation, draft-chen-oauth-rar-agent-extensions, draft-chen-oauth-scope-agent-extensions, and draft-jiang-oauth-intent-admission. This is not improper, but the WG should read the document as partly an advocacy framework for a specific slate of proposals rather than a neutral survey, and should be alert to gaps being framed in ways that presuppose a particular solution.
High-Level Summary
For a VP audience, the document makes one argument in three moves.
Move one: OAuth's consent model assumes a human is watching. OAuth was designed for a user sitting at a browser, approving a fixed list of permissions once, for an application that will use them right away. Every part of that is false for agents: the user states a goal on Monday, the agent acts on Thursday, and neither the user nor the authorization server can reconstruct why a given permission is being asked for at the moment it's asked. The draft calls this "context collapse," and identifies its sharpest form as silent execution within a valid grant — the agent does something the user would not have approved, using authority the user genuinely did grant. Nothing in OAuth is violated; the token is valid; the outcome is still wrong.
Move two: multi-agent chains have no verifiable provenance. When an insurance intake agent hands to a verification agent, which hands to adjudication, which hands to payment, each hop today gets a fresh token and the resource server at the end sees only its immediate caller. There is no interoperable way to prove the chain traces back to the customer's authorization, no way to guarantee each hop holds strictly less authority than the one before, and no standard field to carry a shared identifier (the claim_id) so an auditor can reconstruct the transaction afterward. The draft correctly identifies this as the blocker for enterprise adoption — not security in the abstract, but auditability, which is what actually gates B2B deployment.
Move three: when something goes wrong, there is no off switch. RFC 7009 revokes one token per call. If a laptop is compromised at 2am and an automated responder needs to cut off every token that user holds across every client, it must enumerate and revoke them one at a time, and the draft argues the window this opens is unacceptable when the tokens are held by autonomous software rather than a human who has gone to bed.
The document then grades four existing extensions against agent requirements — RAR (right structure, but assumes values are known upfront), DPoP (right idea, but assumes a stable long-lived client with a persistent key, not ephemeral sub-agents), CIMD (good for onboarding, but its vocabulary describes applications rather than agents), and transaction tokens (good within one trust domain, no attenuation or cross-organizational enforcement) — and sketches evolution directions for each: "bounded capabilities" for RAR, "multi-hop DPoP," agent metadata fields, and an "attenuating agent token."
What it asks for: nothing concrete. It ends with a six-item gap list and an invitation to the WG. There is no prioritization, no proposed work split, and no statement of what OAuth should decline to solve.
Interoperability and Necessity
A use-cases document produces no implementations, so the test is different: does it usefully constrain the work that follows? Partially, and it falls short in one specific way that matters more than any of its technical errors.
The missing "out of scope" determination
The most valuable output of a WG use-cases document is deciding what the working group will not try to solve. OAuth core's decision to leave user authentication out of scope is the reason authorization servers could innovate on passkeys without breaking a single client. This draft makes no such determination anywhere in 39 pages, and at least two of its use cases are ones OAuth should explicitly decline:
- UC4 (agent accessing OS resources) — the draft's own gap analysis concedes an "architectural mismatch": OAuth is network-based and has no notion of process-level authorization, and
sudo/UAC assume a responsible human. Having correctly diagnosed that OAuth is the wrong tool, the draft keeps the use case anyway. It should be dropped, or retained explicitly as an example of a boundary. - UC5 (first connection to a service with no authorization front channel) — the scenario is a REST service with no authorization server at all, and the requirement is a way for a user to mint a scoped, revocable credential without a browser. If there is no authorization server, there is no OAuth. This is a real developer pain point, and it is not an OAuth WG problem; the draft's own framing ("the env-var API key workaround is itself the anti-pattern") is a product observation, not a protocol gap.
Removing these two would strengthen the document, not weaken it. A use-cases draft that draws boundaries is far more useful to a working group than one that catalogs everything.
Requirements are aspirations, not criteria
Each use case ends with a requirements list, but the requirements are written as goals ("verifiable delegation chain," "task-scoped revocation," "execution-layer evidence") rather than as criteria a proposed mechanism could be tested against. Two competing proposals could both claim to satisfy every requirement in the document. If this is to serve as the WG's yardstick, requirements need to be falsifiable — what must an implementation demonstrate to have satisfied "verifiable delegation provenance"?
Overlap and redundancy among use cases
UC6 (business process chain), UC8 (DNSSEC DS maintenance), UC9 (managed services), and UC10 (cross-org delegation) produce nearly identical gap lists: multi-hop attenuated delegation, a shared audit identifier, task-scoped revocation. UC8 is the most interesting of the four because it is a concrete existing protocol with a real RRR trust model rather than a hypothetical, and it should probably be kept as the exemplar with the others compressed into a short subsection noting the same requirements recur. Eleven use cases producing six gaps is a signal that the use cases are over-enumerated relative to their analytical yield.
Substantive accuracy problems
Several gaps are stated as absent when the mechanism exists. Each of these will draw a correction on the list, and fixing them proactively would help the document's credibility considerably:
"No standard for how to nest these claims" (§4.2.2) is incorrect. RFC 8693 §4.1 explicitly defines nested
actclaims: "A chain of delegation can be expressed by nesting oneactclaim within another. The outermostactclaim represents the current actor while nestedactclaims represent prior actors," with a worked example. The real gap is the sentence immediately following in RFC 8693: "the consumer of a token MUST only consider the token's top-level claims and the party identified as the current actor... Prior actors identified by any nestedactclaims are informational only and are not to be considered in access control decisions." That is a much sharper and more defensible finding — the representation exists but is explicitly non-authoritative for access control — and the draft would be stronger for stating it that way.The "Global Token Revocation API" called for in UC11 already has a draft.
draft-parecki-oauth-global-token-revocation-06does precisely what §3.3 and UC11 ask for: a single request to an authorization server to revoke all of a user's tokens and force re-authentication, motivated by exactly the security-incident scenario in UC11. It is not cited. Neither is the authors' owndraft-chen-oauth-agent-revocation. The mass revocation gap needs to be re-scoped around what the existing draft doesn't cover — most plausibly task-scoped and group-scoped revocation, which are genuinely unaddressed — rather than presented as an unaddressed gap."Lack of a standardized interactive channel" (§6, item 2) overlooks two existing mechanisms. The gap is stated as no way for an agent to "pause and securely ask the user for an intermediate decision or to respond to a real-time authorization challenge from a resource server." RFC 9470 (OAuth 2.0 Step-up Authentication Challenge) defines exactly the resource-server-initiated challenge, and
draft-ietf-oauth-first-party-appsdefines an authorization challenge endpoint for non-redirect interaction. Neither is cited anywhere in the document. This is the most consequential omission, because "continuous dialogue" is the draft's first and most emphasized gap, and the WG has adopted work in the space.§5.4 treats Transaction Tokens as folklore rather than a specification. It describes the concept as "a common architectural pattern, often implemented using JWTs" — but
draft-ietf-oauth-transaction-tokens-11is an adopted WG draft, and it appears in the draft's own reference list while going uncited in the section that analyzes it. The analysis should engage with what the specification actually says.§5.3 conflates two different specifications — "CIMD (and the broader concept of dynamic client registration)" merges the Client ID Metadata Document with RFC 7591 Dynamic Client Registration, which are different mechanisms with different trust models. The reference name is also wrong: the WG draft is
draft-ietf-oauth-client-id-metadata-document, notdraft-ietf-oauth-client-metadata.RFC 2119 and RFC 8174 are listed as references with no Requirements Notation section, and there is a single stray uppercase "MUST" in §4.2.3 ("it MUST require one access token at a time") in what is plainly descriptive prose. Either drop the two references or add the boilerplate; lowercase the stray keyword.
Appendix A duplicates §10. The reference list appears twice, in two different formats. This looks like a tooling artifact and should be removed.
Mechanisms worth carrying forward
Not everything here should be cut. Three items are genuinely useful contributions to the WG's framing:
- The grant layer / execution layer distinction is a clean way to explain why RAR and DPoP, however good, don't answer "did the human approve this transaction." It is the document's best idea.
- The Data Subject role is a real structural gap with no current answer.
- Agent-operable consent surfaces (§7) — the observation that an agent can click "Approve" itself, so a click is no longer evidence of human consent — is the sharpest security point in the document and connects directly to the WG's cross-device security work. It is buried as the last bullet of the Security Considerations and deserves promotion.
Writing Quality
Better than average for a document with this many first-time IETF authors. The structure is disciplined and parallel — every use case follows scenario / requirements / what works / what's missing, which makes the document navigable despite its length — and the ASCII diagrams in §3 do real explanatory work rather than decorating.
The register drifts, though. Passages read as advocacy rather than specification: "god tokens," "This is a major architectural gap," and most conspicuously the closing bullet of §7, which switches to second person and consumer-product phrasing: "We need approval methods that agents can't fake, like a confirmation on your phone or a fingerprint scan." The underlying point is the best one in the section; the phrasing undercuts it.
Specific defects:
- The abstract does not stand alone. It opens "This document provides a systematic analysis of these emerging agent-based use cases" — "these" has no antecedent, because the abstract is the first thing a reader sees. This is a one-line fix and it is the first thing anyone will notice.
- A garbled sentence in §7, in the "Durable Record of the Authorization Grant" bullet: "...the exact scopes granted, the time of the grant, and the policy version in effect, which disappear when a grant ends, this event record must persist to provide..." — a comma splice wrapped around a dangling relative clause. Needs to be two sentences.
- Malformed markup at §4.2.4:
- **What's Missing (The Gap)— unclosed emphasis and missing colon, inconsistent with the*What's Missing (The Gap):*used in every other use case. - Missing space in the acknowledgements: "morganLR,Eric Leleu."
- Inconsistent contributor naming — the acknowledgements mix full names, a handle ("morganLR"), a lowercase username ("niyuan"), and a bare first name ("Jeff").
Length and trajectory
At 39 pages this is longer than most Informational use-cases documents need to be, and the trajectory is the more useful signal. Between -02 and -03 the document grew roughly 15% (1,960 to 2,240 lines, ~31 to ~39 pages) while the use case count stayed at 11 and the category structure stayed identical. The growth went into the new §3 visual overview and the grant-layer/execution-layer framing threaded through existing use cases — which is to say the additions were the good part. But a document that expands on each revision without converging on a shorter set of prioritized gaps is a document the working group will find harder to adopt each cycle, not easier. The six gaps in §6 could be argued in roughly half the pages by compressing UC6/9/10 into UC8's analysis and cutting UC4 and UC5.
Background of the Authors
Five authors, with a wide spread of IETF experience — one very senior participant and three with no published RFCs.
Jiankang Yao (CNNIC) — 8 RFCs, and the most credentialed author on the document. Publications span internationalized email (RFC 5336, 6531, 6856, 9755) and registry provisioning (RFC 8543, 8544, 9095), with RFC 6531 cited by 14 other RFCs. Served on the IAB from 2020 to 2024, currently chairs the Technology Deep Dive group and reviews for ART, and holds ICANN ccNSO council and SSAC positions. Participating since IETF 62 (2005), with prior WG chair and secretary experience. 5 active drafts. His presence explains UC8 (DNSSEC DS automation), which is the most technically grounded use case in the document.
Meiling Chen (China Mobile) — 2 RFCs, both DOTS telemetry (RFC 9244, 9387). Extraordinarily prolific as a draft author: 17 active and 42 expired Internet-Drafts across more than a dozen working groups, spanning DDoS signaling, remote attestation, secure routing, post-quantum cryptography, and now agent authorization. Two drafts have reached WG adoption. The ratio — roughly 59 non-replaced drafts to 2 RFCs — indicates a contributor who generates a high volume of early-stage exploratory proposals rather than one who shepherds documents to publication. Six of the active drafts are in the OAuth agent cluster, including the companion proposals this document's gap list feeds.
Peter Chunchi Liu (Huawei) — 0 RFCs, 15 active and 9+ expired drafts. Secretary of the SATP working group. Four of his active drafts carry the draft-ietf- prefix (adopted). His footprint has pivoted from network path validation and NASR (now largely expired) toward AI agent identity, delegation, and attestation, with 5 OAuth-area drafts and 5 on agent identity across WIMSE and related groups. Sustained current engagement, no publication track record yet.
Yuning Jiang (no affiliation listed in the draft header; h-partners.com email) — 0 RFCs, 5 active drafts, all in the agent identity and authorization space: this document, song-oauth-ai-agent-collaborate-authz, jiang-oauth-intent-admission, jiang-wimse-heterogeneous-credential, and jiang-intent-security, plus 3–4 expired drafts in RATS and SPICE. Active but entirely at the individual-draft stage. The missing affiliation on the front page should be filled in.
Jia Chen (China Mobile) — 0 RFCs, 2 active drafts (this one and chen-oauth-scope-agent-extensions), no expired drafts. New to the IETF.
Summary: one long-tenured author with real standing (Yao), one very high-volume proposer with a thin publication record (M. Chen), two active mid-stage participants (Liu, Jiang), and one newcomer (J. Chen). The team is well positioned to enumerate use cases but has limited collective experience converging a document to publication — which matches the trajectory concern above.
Suggested Feedback
Ordered roughly by what would most improve the document:
Add an explicit "out of scope" section, and put UC4 and UC5 in it. The draft already concedes UC4 is an architectural mismatch and UC5 describes a service with no authorization server. A use-cases document that tells the WG what not to work on is worth more than one that catalogs everything. This is the single highest-value change.
Correct the RFC 8693 claim and sharpen it. Nested
actclaims are defined in §4.1. The real finding — that RFC 8693 declares prior actors "informational only and not to be considered in access control decisions" — is a stronger argument for the same conclusion.Engage with the existing work on the "continuous dialogue" gap. RFC 9470 and
draft-ietf-oauth-first-party-appsboth address resource-server-initiated and non-redirect user interaction. The gap needs to be restated as what those mechanisms don't cover for agents, not as an absence.Re-scope the mass revocation gap around
draft-parecki-oauth-global-token-revocation. Per-user global revocation has a draft at -06. Task-scoped and group-scoped revocation appear genuinely unaddressed and are the part worth keeping.Cite
draft-ietf-oauth-transaction-tokensin §5.4 rather than describing transaction tokens as an informal pattern, and separate CIMD from RFC 7591 in §5.3 (fixing the reference name todraft-ietf-oauth-client-id-metadata-document).Promote the Data Subject role and the agent-operable-consent observation. These are the two most original points in the document and both are currently buried — the first in a terminology list, the second in the last bullet of Security Considerations.
Make requirements testable. Rewrite each requirements list as criteria a candidate mechanism could be evaluated against, and prioritize them. Without this, the document cannot serve as a yardstick for the proposals that cite it.
Compress rather than expand in -04. Fold UC6, UC9, and UC10 into UC8's analysis, note that the same requirements recur, and target a shorter document. The -02→-03 growth added good material, but continued growth will make adoption harder.
Editorial pass: fix the abstract's dangling "these"; repair the §7 comma splice in the durable-record bullet; fix the §4.2.4 list marker; add Yuning Jiang's affiliation; remove the duplicate reference list in Appendix A; either add a Requirements Notation section or drop RFC 2119/8174 and lowercase the stray "MUST" in §4.2.3; normalize the acknowledgements to consistent full names.
A process question worth raising on the list: whether the WG wants an Informational use-cases RFC at all for this space, or whether the requirements are better captured as a short problem-statement section inside whichever solution documents get adopted. Use-cases RFCs age quickly, and this one is expanding rather than converging. If the WG does want one, it should say what the document must decide — priority ordering and scope boundaries — rather than adopting a catalog.