Review · October 26, 2026
Seven Strings to Say What Kind of Thing Is Calling
First, what changed in -02
Nothing. The document history for -02 says "No changes" and a diff against -01 confirms it. This revision was posted on October 16, one day before -01 would have expired. It's a keepalive, not an update.
That matters because six of Karl McGuinness's individual drafts now cite this document normatively, including the three-part actor suite reviewed earlier this month. The draft hasn't moved in six months, but the weight resting on it has grown.
The problem it's solving
A resource server looking at a token today can tell who the subject is and which client requested it. It can't tell whether the subject is a person, a service account, or an AI agent, or whether the client is a mobile app, a backend, or an agent runtime. Each authorization server encodes that in its own custom claims, if at all.
The draft proposes one shared vocabulary for "what kind of thing is this," carried wherever OAuth already carries identity: JWT access tokens, ID tokens, JWT assertions, Transaction Tokens, introspection responses, dynamic client registration, and AS metadata.
How it fits in the working group
In charter. It's an extension that registers claims and metadata parameters against existing RFCs. No new endpoints, no new flows, no new roles.
The dependency list is wide. It touches RFC 9068, Dynamic Client Registration, Token Introspection, AS Metadata, Token Exchange, RFC 7523 JWT assertions, OpenID Connect Core, and the adopted Transaction Tokens draft. Transaction Tokens is a normative reference, so this can't publish until that does.
It does add one new responsibility to the AS: vouching for a classification. The RS is told to make access decisions on that classification. How the AS establishes it is out of scope. More on that below.
What the mechanism actually does
The core is small:
- Seven registered values:
user,device,native_app,web_app,browser_app,service,ai_agent. Three of those are the client types already in Section 2.1 of RFC 6749, renamed. - Two claims.
client_profileclassifies the client inclient_id.sub_profileclassifies the subject insub. - Values are space-delimited, multi-valued, case-insensitive, unordered, with no hierarchy. "service ai_agent" is legal and means both, independently.
sub_profilecan also appear insideactnodes from RFC 8693 to classify each hop in a delegation chain.- A new IANA registry with a "Usage Location" column saying whether each value is valid as a client profile, subject profile, or actor profile.
- One AS metadata parameter,
entity_profiles_supported, withclient,subject, andactorarrays. - One DCR parameter. The client proposes a
client_profile; the AS is authoritative and returns what it actually assigned.
The rest of the 36 pages is behavior sections for client, AS, and RS, plus 13 security subsections and 5 privacy subsections.
So what does the RS do with it? Section 10 says use it for access control, and apply default-deny when the AS advertises support for a usage location but the claim is missing from the token. The RS has to fetch AS metadata to know whether a missing claim is meaningful.
Significant interoperability concerns
The values are unverifiable and the spec says so. Section 3.1.7 admits "it can be difficult to objectively verify that an entity truly qualifies as an AI agent." Section 6 says verification of client_profile is out of scope. Section 9.1 then says the AS MUST verify it. In an open dynamic registration world, the kind MCP is pushing toward, the AS has nothing to verify against. The value is self-asserted by the client, the AS passes it through, and Section 10 tells the RS to make access decisions on it. That's the chain the draft builds, and Section 12.2 warns against it without resolving it.
The authors' own examples disagree with each other. Appendix A classifies a smart TV app in the Device Authorization flow as native_app, not device. The "AI acting as itself" row gives the client "service ai_agent" and the subject "ai_agent service" for the same entity. A web app calling a downstream API gets "service web_app". Section 3.3 says conflicting profiles SHOULD be avoided, then immediately blesses "user ai_agent" for a digital employee. If the authors can't apply the vocabulary consistently in a one-page table, two independent ASes won't either.
Multi-value with no hierarchy breaks the "most restrictive" rule. Section 10.1 says when policies conflict, apply the most restrictive. But Section 3.3 says no value is a subtype of another and each is interpreted independently. An RS seeing "service ai_agent" has to decide whether that's a service with agent features or an agent with service privileges, and the spec says both readings are wrong because there's no relationship. In practice every RS will pick one.
client_profile overlaps OpenID Connect's application_type. OIDC Dynamic Client Registration already has application_type with web and native. The draft doesn't mention it. Any OIDC provider implementing this ends up with two parameters describing the same thing.
Default-deny couples RS behavior to AS metadata. Section 10 says penalize missing claims only for usage locations the AS advertised. Section 12.8 then says consider backward compatibility and accept requests without them. An RS reading both gets no clear instruction.
Mechanisms that could be removed
Actor profiles. Section 11 says the rules for constructing act chains and assigning actor profiles "are expected to be defined in a separate specification." That specification exists: it's draft-mcguinness-oauth-actor-profile. The actor usage location, the actor metadata array, and three security subsections could move there. This draft would shrink by a quarter.
The behavior sections. Sections 8 through 10 are almost entirely MAY and SHOULD with policy left to the deployment. The normative content reduces to: values must be syntactically valid and registered or private, the AS is authoritative over client_profile, the client never influences sub_profile, and introspection wins over the JWT on mismatch. That's a page.
Repeated requirements. "MUST NOT preserve invalid values" appears in Sections 3.3 and 9.6. "MUST NOT infer meaning beyond the definition" appears in 9.6, 10, and 12.4. The pseudo-code, the OIDC claims parameter example, and the error-message text that ends by saying error messages aren't standardized can all go.
The prose definitions. The user definition is a paragraph of assumptions about human cognition ("It is assumed that users can make informed consent decisions"). The ai_agent definition describes machine learning. None of it is testable. The registry's one-line descriptions are better than the Section 3.1 essays.
Writing quality
Fluent and in the IETF register. Sentences are well formed. The structure is conventional.
It's too long for what it defines. Two claims, seven strings, and a registry don't need 36 pages. Section 8.1 argues that "clients are incentivized to declare accurate values," which is advocacy, not specification. The security section uses lowercase "must not" throughout while the rest of the document uses BCP 14 capitals, so it's unclear which of those are requirements. Em-dashes appear in the abstract and body.
The best-written part is Section 3.3's acme_verified_robot example explaining why there's no hierarchy. It's concrete and it would prevent a real mistake.
Author background
Datatracker was unreachable from this session, so these are from prior reviews and local draft files.
Sreyantha Chary Mora (Microsoft) is the lead author. This is the only draft under his name in the local OAuth draft corpus. He appears to be a new IETF author.
Pamela Dingle (Microsoft) had zero RFCs and three active drafts as of August 2026: SPICE glue-id, SCIM agent resource, and this one. Well established in the identity industry through the OpenID Foundation, newer as an IETF author.
Karl McGuinness (Independent) had zero RFCs, sixteen active drafts, and two expired as of mid-October 2026. He joined this document at -01 and has since made it a normative dependency of six of his own drafts.
Questions worth raising
- Is a classification vocabulary a protocol element? The only interoperable part is the seven strings. Everything else is local policy. If the WG wants this, a four-page document registering the claims and the values would do the same job.
- How does
client_profilerelate to OIDCapplication_type? Either reconcile them or explain why a second parameter is needed. - Should actor profiles move to the actor profile draft? This document defers the rules to it already.
- What does an AS do with
client_profilewhen it can't verify it? The draft says MUST verify and verification is out of scope in adjacent sections. Pick one, and tell the RS what the claim's provenance actually is. - Is there WG interest, or only dependent-draft interest? No content has changed since April. The reason to adopt now is that other individual drafts need it, and that's a reason to ask whether those drafts should inline three strings instead.
For the delegation half of this proposal, read draft-mcguinness-oauth-actor-profile. For the client type definitions this draft renames, read Section 2.1 of RFC 6749.