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

Review · November 10, 2026

Scenario-Based Service Discovery and Authorization Protocol in Agent Environments

draft-liu-oauth-scenario-discovery-authz -00View on Datatracker →

An agent lands on an application it has never seen. It doesn't know what the application lets agents do, which API or MCP server to call, or how to get a token for the user.

This draft answers all three with one JSON file at /.well-known/agent-capabilities. The application lists the "flows" it opens to agents, the surfaces that reach them, and which of four OAuth "establishment modes" its AS accepts.

The discovery document is the new part. The other 30 pages profile existing OAuth work into a full agent authorization stack.

How it fits in the working group

The draft says it "defines no new authorization protocol." That's true only in the sense that it adds no grant types or endpoints at the AS.

What it does add:

  • A new well-known URI and a capability document with seven top-level fields
  • One new Protected Resource Metadata parameter, agent_capabilities_uri
  • Three new IANA registries: document fields, establishment mode identifiers and surface types
  • A new role, the Host Container, which is the phone OS or browser the agent runs in
  • A browser_session_endpoint where the host trades the agent's token for a web session cookie
  • A new use of ID-JAG as the token the agent presents to the resource server

Normative dependencies are nine OAuth RFCs, three working group drafts (ID-JAG, Deferred Token Response and SPIFFE client authentication) and the author's own individual draft, Authorization Schemas. All runtime enforcement and error handling is delegated to that last one, so this draft can't progress ahead of it.

Is this in charter? The authorization pieces are. A description of what an application offers to agents is application-layer service discovery. That's closer to RFC 9727 api-catalog, MCP tool listings and A2A Agent Cards than to anything the OAuth WG has published.

It also overlaps with adopted work. RAR Metadata and Error Remediation already advertises authorization_details types and tells a client what to request after a refusal.

What the mechanism actually does

  1. The agent fetches https://shop.example/.well-known/agent-capabilities. A native app also ships a copy in its package, so a host-side agent can read it offline.
  2. The document lists surfaces (http_api, mcp, web, native_app) and flows such as compare-and-buy. Each flow names its operations, its schema types and the surfaces that reach it. An operation no flow lists is closed, and the RS must refuse it.
  3. The agent matches the user's instruction to a flow. The matching algorithm is out of scope.
  4. The agent picks an establishment mode. Mode A is the authorization code flow with PAR, PKCE and RAR. Modes B1, B2 and B3 are Token Exchange requests for an ID-JAG, with user interaction before the request, after it through Deferred Token Response, or not at all.
  5. The AS issues a DPoP-bound JWT with the approved boundary in authorization_details, for example a 200 USD purchase limit.
  6. The agent calls the API or MCP server. When a call exceeds the boundary, the RS returns insufficient_authorization with a schema_profile_uri. The agent goes back to the AS for a narrow "remediation token", retries the one call, then goes back to the original token.
  7. The host container can also carry the session into the app's web site or native app, without the agent seeing the cookie.

Significant interoperability concerns

The Mode B token isn't an ID-JAG. ID-JAG is a grant. An IdP issues it, its aud is the Resource AS, and the client redeems it there for an access token. Here the application's own AS issues it with aud set to the resource, and the agent sends it straight to the API as a DPoP-bound bearer token. The token response says token_type is N_A, which means "this is not an access token." ID-JAG's typ is oauth-id-jag+jwt specifically so a resource server won't accept one. What the draft describes is Token Exchange returning an access token. It should say that.

Nobody is defined as the issuer of the identity assertion. In B1 the user authenticates "with a host-native credential, such as a passkey," and "the resulting identity assertion" becomes the subject_token, typed as an ID Token. A passkey doesn't produce an ID Token. ID-JAG requires the subject_token to be an assertion the AS itself issued to this client. The draft never says which party mints it or why the shop's AS would trust it. In B3 the agent "already holds" one, which contradicts the claim that the agent holds exactly one credential.

Mode B2 sends Token Exchange without a subject_token. RFC 8693 makes that parameter required. The AS then "obtains the identity assertion" on its own after the user approves. How the AS knows which user to ask is not specified. There's no login_hint or equivalent.

The agent "cannot pre-register," and the draft never says how it becomes a client. The examples use a URL as client_id, which looks like Client ID Metadata Document, but that draft isn't cited. The one client authentication method described is SPIFFE, with the trust bundle obtained "out of band." That is a prior relationship.

Operation names have no shared vocabulary. ecommerce.purchase belongs to "the value space of the schema types." The schemas draft defines no normative schema and its registry starts empty. Two shops will name the same operation differently. The agent ends up reading the free-text description, which an LLM can already do with an MCP tool list.

A public cached document becomes access control. The RS MUST NOT execute an operation that no advertised flow lists. The well-known file SHOULD be static content behind a cache. Remove a flow and agents keep requesting it until the cache expires, while the RS has to know the current flow list on every call. The draft also says a stale packaged copy keeps a flow "open only until the first refresh," which conflicts with the same paragraph's MUST NOT.

Per-flow mode restrictions can't be enforced. A flow can exclude silent establishment, as the gift card example does. The RS "MUST NOT accept or reject operations based on any inference about how the token was issued," and the token doesn't record the mode. Only the AS could enforce it, and the AS isn't told which flow a request is for.

Mode A's return path is undefined. The code reaches the agent by "a front-channel redirect through the host container, or a polling or pushed arrangement." For a cloud agent, that sentence is the whole problem. Mode A is the mandatory baseline, yet establishment_modes_supported is OPTIONAL.

"Session" has no protocol object. The AS MUST revoke remediation tokens when "the delegated session ends." No session identifier appears in any token. The tokens are short-lived JWTs with no revocation mechanism defined. The agent juggles two tokens and must pick the right one per call.

One surface per type. type MUST be unique in invocation_surfaces_supported. An application with two APIs or two MCP servers can't describe itself.

The web and native handoffs are deployment-defined. The browser_session_endpoint request and response are listed as an open issue. The native IPC channel is "the host platform's secure credential passing." Two implementations can't interoperate on either.

Mechanisms that could be removed

Section 5, all four modes. Which grants an AS supports belongs in AS metadata, where grant_types_supported already lives. The draft says the mode identifiers exist because three modes share a grant type. Fixing the ID-JAG misuse removes most of that need. Deferred Token Response already has its own discovery.

The web and native presentation surfaces. They carry the most new security surface and the least specification. Token-to-cookie exchange is its own draft if anyone wants it.

The service block. PRM already has resource_name, resource_policy_uri and resource_tos_uri.

The remediation rules in Section 7. They restate the schemas draft and add agent behavior requirements ("MUST suspend the affected task") that no server can observe.

Section 4.3. Half a page on Web Bot Auth that ends with "this specification requires neither."

Two of the three registries. A new draft with four surface types and four modes doesn't need Specification Required registries yet.

What's left is a small document: a list of API and MCP endpoints, the flows they serve, and a pointer to the AS. At that point the question is whether it should be a new well-known file or a few more fields in Protected Resource Metadata.

Writing quality

The prose is fluent and confident. BCP 14 keywords are used carefully in most places, and the open issues appendix is candid about five things that aren't done.

It's too long. 46 pages for one JSON document and a profile. The same sentences recur: "the approved schema instance, not the issuance path, is the authorization" appears three times, the Android explicit intent advice three times, the 401 fallback explanation twice in adjacent sections.

"Scenaric" isn't an English word. It's in a field name, scenaric_flows_supported, so it will be hard to fix later. The draft's own IANA section slips and calls them "scenario flows."

The text has rhetorical lines that don't belong in a spec. "Discovery is the protocol's front door." "The protocol's credential story is one sentence." "Is a signal, not an attack."

There are leftover edits. flow_id is unique within "this metadata document," though the draft insists the capability document is not metadata. Applications "MUST NOT advertise a grant type," and the document has no grant type field. The appendix has the user set a 5000 USD ceiling, the agent request 200 USD, and then hold a 480 USD fare "within bounds."

Background of the Authors

Datatracker couldn't be reached for this review. The numbers below come from our earlier 2026 reviews.

  • Dapeng Liu (Alibaba): 3 RFCs from 2014 and 2015 in distributed mobility management, around 10 active drafts and 80+ expired ones. Not a new author. He is writing a lot right now: this is his fourth OAuth agent draft we've reviewed this cycle, alongside Authorization Schemas, authorization evidence and delegation chaining. He is also a co-author of the adopted Deferred Token Response draft, which Mode B2 depends on.

Liu is the sole author. Hongru Zhu, Hui Xue and Suresh Krishnan are acknowledged.

The drafts are starting to stack. This one depends normatively on Authorization Schemas, whose -00 feedback is still open. Feedback is best aimed at the bottom of the stack first, then at the ID-JAG misuse here.