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

Review · November 10, 2026

Transaction Tokens for Agents Adds an Agent Chain Claim to Txn-Tokens

draft-araut-oauth-transaction-tokens-for-agents -03View on Datatracker →

The problem it's solving

Transaction Tokens carry user identity and transaction context through an internal call chain. They say nothing about whether an AI agent was involved.

That's the gap this draft targets. A database three hops deep can't tell a direct human API call from a request that passed through two autonomous agents. Audit logs record the operation but not the agent path that caused it.

The draft adds one new claim, agentic_ctx, to the Txn-Token. It names the agent that started the chain, the agent at the current hop, and how many agent hops have happened.

How it fits in the working group

It's a claim extension to an adopted working group draft, so it's in charter. It defines no new grant types, token types, or endpoints.

The dependency list is short: Txn-Tokens, RFC 8693 Token Exchange for the act claim, RFC 9068 for JWT access tokens, and RFC 6749.

It introduces one new role, the Agent Attribute Store. This is a registry inside the trust domain that tells the Transaction Token Service (TTS) which workloads are agents and what assurance level each has. Resource servers can also query it. The draft defines no protocol for it.

So does it line up with the current Txn-Tokens draft? Not quite, in three places.

  • The whole mechanism runs on replacement tokens. Txn-Tokens removed its replacement section, and what's left is Section 13.15, which warns that replacing Txn-Tokens "potentially negates the primary purpose of having Txn-Tokens." This draft requires a replacement at every agent hop and links to a section that no longer exists.
  • It says req_wl is appended with a comma separator "per the base specification." Txn-Tokens -11 defines req_wl as a single string identifying the requesting workload.
  • Section 5.13 describes a Txn-Token leaving the trust domain with attr_store_ref stripped. Txn-Tokens says a Txn-Token MUST NOT be accepted outside its trust domain.

What the mechanism actually does

At initial issuance, the API gateway sends the external access token to the TTS as usual. The TTS decides whether the client is an agent, by Agent Attribute Store lookup, by the presence of act, or by local convention. If it is, the TTS copies any act claim into the Txn-Token and adds agentic_ctx.

agentic_ctx contains:

  • originator.id: the agent that started the chain. Never changes.
  • proximate_agent.id: the agent at the current hop. Updated on each replacement.
  • chain_metadata.hop_count: starts at 1 and increments on each agent replacement.
  • chain_metadata.min_assurance_level: optional. The lowest assurance level of any agent so far.
  • session_id on both agent objects: optional, self-asserted, for log correlation only.
  • attr_store_ref: optional URI of the Agent Attribute Store.

When an internal agent needs to call downstream, it asks the TTS for a replacement Txn-Token. The TTS keeps sub, txn, aud, act, and originator fixed, updates proximate_agent, increments hop_count, and lowers min_assurance_level if the new agent's level is lower. Non-agent workloads carry the claim forward untouched.

The assurance rule is the interesting part. A low-trust third-party agent can't raise its standing by routing through a high-trust internal agent. The draft calls this preventing "identity laundering."

The token only shows the first and current agents. The full path lives in a TTS audit log, keyed by txn.

Significant interoperability concerns

Txn-Tokens live inside one trust domain with one TTS. That lowers the bar. Deployment-defined assurance values are fine when one operator controls both the TTS and the resource servers. The claim structure is what needs to be standard, and that part is specified clearly.

The problems are elsewhere.

Agent identification is undefined. The draft says "the specific mechanism for agent identification is deployment-defined." That's acceptable for a single domain. The procedure it does give is confusing, though. It tells the TTS to extract client_id from "the presented access token" on replacement requests, where the subject token is a Txn-Token and the requester is an internal workload.

The act claim means something different here. In RFC 8693 the top-level act is the current actor and nested act claims are prior actors. This draft freezes act as the original agent forever. A resource server applying RFC 8693 semantics will read it wrong. originator.id then duplicates act.sub, which the draft admits is "for convenience." The draft also never explains why nested act wasn't used for the chain, which is the approach draft-mw-oauth-actor-chain takes.

session_id has no way to reach the TTS. The field is supplied by the agent's infrastructure, but no request parameter is defined. For a third-party agent, the API gateway makes the Txn-Token request, so it's unclear how that agent's session ID gets there at all.

Replacement is optional, so the claims may be stale. Agents SHOULD request replacement. Section 5.14 concedes that otherwise hop_count is only a lower bound. The draft says deployments MUST treat unreplaced tokens as non-compliant, but a resource server can't detect one.

Normative language conflicts. Section 4.2.3 says the TTS MUST enforce a maximum hop count. Section 5.12 says SHOULD.

No IANA Considerations. agentic_ctx isn't registered as a JWT claim.

Mechanisms that could be removed

Roughly half the document doesn't affect what goes on the wire.

  • Token Exchange Isolation (3.7, 5.16). MUST-level rules that the LLM can't call the token exchange or see credentials. This is sound advice, but it's agent runtime architecture. The draft itself says a Txn-Token carries no evidence it was followed.
  • TTS Audit Logging (3.8). A list of log fields. Operational guidance, not protocol.
  • Agent Attribute Store contents (3.6.2). The store's interface is out of scope, yet the draft specifies its fields, including posture, provenance, and attestation data.
  • attr_store_ref. A URI pointing at an endpoint with no defined protocol can't be used interoperably.
  • session_id. Self-asserted, unverifiable, and forbidden from authorization decisions. A TTS log entry would do the same job.
  • Sections 5.1 through 5.8. Mostly generic statements like "implementations MUST verify signatures" that restate the base spec.

What's left is a small, useful core: a claim with originator, current agent, hop count, and a floor on assurance level.

Writing quality

The structure and vocabulary are standard IETF. The prose is mostly clear.

It's too long for what it defines. One claim with six fields takes 40 pages. The same rules appear three times: in 3.3.1, again in 4.2.1, and again in the security considerations. The page count grew by about a quarter since -02.

There are signs of fast revision. -03 renamed "Agent Registry" to "Agent Attribute Store" and missed three places. There are two reference sections. The Txn-Tokens reference points to the pre-adoption individual draft. Section 3.3 says "see Section 3.3 below." A few sentences are ungrammatical, such as "utilize Large Language Models (LLM)s and plans, reasons,and takes actions." The abstract contains em-dashes and inline URLs.

The biggest missing justification is the choice not to use nested act for the chain. The second is why replacement is acceptable here when the base spec discourages it.

Background of the author

Ashay Raut (Amazon) is the sole author. I couldn't load the Datatracker profile for this review, so RFC and draft counts are unverified. From the documents in hand:

  • Raut is listed in the acknowledgments of the Txn-Tokens working group draft.
  • Raut also authored draft-araut-oauth-transactiontokens-bcp in July 2026, so there are at least two active individual drafts, both about Txn-Tokens.
  • This draft is on its fourth revision since early 2026, which shows sustained effort.

Atul Tulshibagwale, lead editor of Txn-Tokens, is listed as a contributor. That suggests the Txn-Tokens editors are aware of this work.

Bottom line

The core idea is small and reasonable. Internal services do need to know when an agent is in the call path, and the assurance floor is a good rule.

The draft needs to be cut down to the claim definition and the replacement rules. It then needs to reconcile with Txn-Tokens -11 on replacement, req_wl, and trust domain boundaries, and explain its relationship to RFC 8693 act semantics and the actor chain work.