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

Update review · November 22, 2026

Security Principal and Verifier Binding for Agent Communication Protocols (Update -09)

draft-bu-agentproto-security-principal-binding -09View on Datatracker →

What Changed

Revision -09 adds material and removes nothing. The draft grew from 74 pages to 80.

The previous review said the matrix, the status values and the "accepted result" concept would fit in ten pages. This revision goes the other way.

The claim registry, row fields, outcome model and composition rules are unchanged. The additions are two illustrative scenarios, four more negative test cases, and one more pinned test report against someone else's code.

New in This Revision

An HTTP dialog-resume example (Section 8.1). A gateway decides whether a service agent, acting as an OAuth client, may resume an existing dialog after a restart or key change. The agent presents a JWT access token bound with DPoP. A separate JWS carries an "application-specific continuity statement" that links the old instance and key to the new one.

The statement binds a long list: tenant, audience, logical identity and incarnation, old and new key thumbprints, predecessor identifier and digest, remaining scope, a sequence number, a relying-party challenge and a validity interval. The gateway commits the new head atomically, so two competing successors cannot both win.

A deployment question (Section 8.2). A coordinator has to decide whether a replacement service instance may rejoin an existing dialog. The section separates three decisions: the principal's grant of authority, the binding of that grant to an instance or key, and the relying party's decision to resume. It points at Section 4.2 of the adopted use cases and requirements draft.

Four new negative cases. All four come from the dialog-resume example. The count is now 66.

An AAE/A2A test report (Section 32). A local harness ran eight case groups against a pinned commit of an A2A extension repository.

Was the Previous Feedback Addressed?

Mostly no.

  • Nothing to interoperate with. Unchanged. The new continuity statement makes this worse. It is a signed object that drives an access decision, and the draft gives no claim names, no example and no issuer. The draft calls it "not a registered claim, a new generally applicable credential format, or a complete OAuth or A2A profile".
  • MUST without a format. Unchanged. The keyword still appears about 50 times.
  • Unstable claim registry. Unchanged. Logical-agent continuity now has two more pages of text and still no claim ID. The open issues list says so directly.
  • Inconsistent vocabulary. Not fixed. Sections 24.3 and 24.5 still say "Specification status: external". Section 32 still says "thirteen claim classes".
  • Mechanisms that could be removed. None were removed. Section 32 gained another pinned-commit test report, which is the kind of content the last review suggested moving to a mailing list post.
  • No complete filled-in row. Partly addressed. Section 8.1 is the first concrete example that names real carriers. It is written as prose inside the principal model section. It is still not a filled-in matrix row, so the draft never demonstrates its own template.
  • Charter fit. Section 8.2 is the first text that ties the draft to the adopted requirements document. That is a step in the right direction.

Does the OAuth Example Hold Up?

The OAuth statements are accurate. DPoP proves possession of a key for one request. It is not client authentication, it does not cover the request body, and it cannot tell two processes holding the same key apart. The draft says all of this.

The DPoP replay cache and the continuity commit are kept separate, which is correct. A retry needs a fresh proof.

The problem is who signs the continuity statement. The draft calls it an "eligible transition authority" and leaves provisioning and status to the deployment.

In OAuth, a client that changes its DPoP key gets a new access token from the authorization server. The client_id and sub are already the stable identifiers across that change. The example needs to say why the gateway cannot key the dialog to those, and why a second signed artifact from a second authority is required.

The draft half-answers this itself. Section 8.1.4 says a complete binding-and-policy implementation "can make the same decision as this mapping" and that "no algorithmic advantage over that baseline follows".

Writing Quality

The same pattern continues. Each revision absorbs a discussion thread and adds its distinctions as new paragraphs.

Section 8.1 spends as many words on what it does not claim as on what it does. It is not a profile, not a format, not conformance evidence, not interop evidence and not working group agreement.

One citation issue: Section 8.2 cites the use cases draft by inline URL to a specific revision instead of a reference entry.

Authors

No change. Songbo Bu remains the sole author.

Questions for the Working Group

  • Is dialog resumption after instance replacement a requirement for the dialog management protocol? Section 8.2 frames the question well in about a page. That page is the part worth discussing on the list.
  • If it is a requirement, does it need a continuity artifact separate from the access token? Or is reissuing a token under the same client identifier enough?
  • The first open issue from -08 still stands: is this a requirements document, a security considerations companion, or text to import elsewhere? Nine revisions in, the draft has not picked one.

The use cases and requirements draft is still the place where this feedback will have an effect. Section 8.2 could be proposed there as a short addition to the recovery requirements.