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

Review · October 26, 2026

Actor Chain Overview -00, A Ten-Page Map of the Actor Chain Protocol

draft-mw-oauth-actor-chain-overview -00View on Datatracker →

The problem it's solving

This draft adds no new protocol. It is the Informational companion to Actor Chain -02, which was split out of the 109-page -01.

The problem is the same one. A request passes through orchestrator A, planner B, and connector C before reaching API D. Token Exchange can record that history in nested act claims. RFC 8693 treats every actor except the current one as informational, and it has no rules for building, trimming, or verifying that history. Actor Chain adds those rules and leaves the AS in charge of issuing tokens.

The overview explains the design in ten pages so reviewers don't have to start with the 63-page protocol document.

How it fits in the working group

As an overview it requests no IANA action and defines nothing normative. Its only normative effect is pointing at the protocol draft. It cites Token Exchange, Protected Resource Metadata, Introspection, DPoP, mTLS, and SD-JWT, plus MCP, A2A, and Kubernetes service accounts as example deployments.

So should the WG want an overview draft at all? The WG rarely adopts standalone overviews for single proposals. This one is most useful as a reading aid while the WG compares Actor Chain with the other delegation-evidence drafts: Actor Proofs, Delegation Chain, and the transaction token and WIMSE variants.

What it describes

Two axes, six profiles. Assurance is either Declared (the AS asserts the history) or Verified (each actor signs its transition request and the AS signs an acceptance). Disclosure is Full, Subset, or Actor-Only. The profile is picked when the workflow starts and stays fixed.

Disclosure and assurance are separate. A Verified token can still show D only C. The draft's payroll example fits this case: an external payroll API authorizes the registered payroll connector and never learns which internal HR agents prepared the request.

Four new claims. actp names the profile, acti is a random workflow ID shared by all tokens in the workflow, act carries the current actor, and actc is the AS-signed acceptance. That acceptance hashes C's proof and B's acceptance, so the chain links by hash without embedding earlier records.

The resource decides how much evidence it needs. D advertises its requirement in Protected Resource Metadata. Evidence arrives in an acte claim, either inline in the token or fetched by introspection. Coverage is current (C's step only), disclosed (original proof and acceptance for each actor visible in act), or path (the full chain, Full disclosure only).

Cross-domain handoff preserves signatures. When the workflow moves to another AS, the accepted state carries over unchanged. The draft is clear that if the destination maps read to documents.read, the earlier signature still covers only read.

It is honest about limits. Under disclosed coverage, the AS can still misstate who was on the path. A disclosed A/C subset proves nothing about whether B sat between them. An actor's exchange-time signature says nothing about who is presenting the token now. MCP's bearer tokens need a sender constraint before any of this means anything. The draft also says plainly: "Live interoperability remains untested."

Significant interoperability concerns

The protocol-level concerns are in the -02 review. The overview makes some of them easier to see.

Six profiles is a lot for three use cases. Table 1 gives one example per disclosure mode. None of them needs Declared assurance. The overview never explains when a deployment would pick Declared over plain RFC 8693, which supports cutting it.

Envelope versions show up as feature flags here too. The acte example uses "v": 1 for current. A few sentences later, disclosed "uses envelope version 2." A reader of the overview won't guess that the version number decides which coverage modes are valid.

The user is still missing. The illustrative workflow begins with orchestrator A already holding a token. sub is described as "the party or context represented by the workflow" and shown as workflow-subject-alias. Nothing shows how a user's authorization becomes the root of the chain. For an agent-delegation proposal, that is the first question reviewers will ask.

Actor key binding gets one clause. Security Considerations depends on "independently authorized ActorID bindings." The overview never says where those come from. The protocol draft doesn't either.

Mechanisms that could be removed

  • Section 10, Design Evolution, and the lineage in the abstract. The SPICE history appears twice in ten pages. It belongs on the mailing list.
  • The MCP, A2A, and Kubernetes framing in Sections 2 and 6. The workflow works without naming protocols. Section 6 mostly lists mismatches (MCP bearer tokens, DPoP vs mTLS) that belong in the protocol draft.
  • The SD-JWT paragraph. It says SD-JWT could be used later and defines nothing. Move it to future work.
  • The overview itself, eventually. If the protocol draft dropped its informative Sections 27 through 35, this document could be its introduction. Two documents with overlapping prose will drift out of sync.

Writing quality

It is short and well organized. The tables in Sections 1.1, 3, and 5 are the clearest explanation of this proposal so far. The "Questions for IETF Review" section is a good move and should be the starting point for discussion.

The prose is compressed past the point of clarity. Sentences like "Current evidence can hide A/B. disclosed uses envelope version 2; each selected signature is checked, and the final acceptance equals actc." assume the reader already knows the protocol draft. Section 2 says the same thing twice in a row: "C signs an actor proof... then signs an acceptance. C signs the request; the AS signs its acceptance." Lowercase claim names start sentences without code formatting, which makes them read like typos.

There are small inconsistencies. The reference to the protocol draft lists four authors and a 4 October date, while this draft has five authors. The Motivation table and the A→B→C→D example use different scenarios for Subset, so the reader has to switch between two mental models.

Background of the authors

Datatracker profiles could not be fetched for this post. Figures are carried over from earlier reviews of this author group.

  • A. Prasad (no affiliation listed): 0 RFCs, around 7 active drafts across OAuth, SPICE, RATS, and WIMSE. A new author writing at a high rate.
  • Ram Krishnan (no affiliation listed, previously JPMorgan Chase): 6 RFCs going back to 2002 in MPLS and networking, around 7 active drafts, recently moved into security work.
  • Diego R. Lopez (Telefonica): 12 RFCs, around 22 active drafts, chairs GREEN. Network operations background, not identity.
  • Srinivasa Addepalli (Aryaka): 0 RFCs, around 7 active drafts that overlap with Prasad's.
  • Suresh Krishnan: dozens of RFCs, former Internet Area AD and IAB member. He is listed on this overview, so he is on both halves of the split.

All five authors are on both documents. Two personal Gmail addresses and no affiliations make it hard to tell which organizations are behind the proposal.

Questions worth raising

  1. Start with the authors' Question 1. Do these use cases need a common contract beyond RFC 8693? If yes, the WG should answer it once for all five delegation-evidence drafts, not separately for each.
  2. Where does the user enter the chain? An overview for agent workflows should show the human at the root.
  3. What is Declared assurance for? Ask for one use case where Declared is the right choice and plain Token Exchange isn't enough.
  4. Is anyone building this? "Live interoperability remains untested" after four revisions in two working groups.

For the normative details, read Actor Chain -02. For the competing inline-proof design, read Actor Proofs.