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

Review · November 16, 2026

The AUDIT BoF Charter, Agent Audit Records Without Saying Who Writes Them

audit-bofView the source →

AUDIT (Agent Use of Delegation and Interaction Traceability) is a working-group-forming BoF. It meets Tuesday November 17, 9:00am to 10:30am Pacific. No solution drafts will be presented. The session is about the charter, so this post reviews the charter.

What problem does the charter say it is solving?

An agent acts for a user across several domains. Each domain keeps its own logs. No single party can reconstruct who started an action, under what authority, and how that authority changed along the way.

The charter wants records that can be collected and correlated across domains. It also sets a high bar: audit information "must be verifiable by a party that trusts neither the agent nor its operator." The charter says that property is what separates audit records from ordinary logs.

What is in scope?

A reference architecture, data representations, and "extensions or adaptations to existing communication protocols for inter-domain correlation." The scope list has four items:

  • Attributing actions to a user, agent instance, or delegation context
  • A record format for delegation chains, "authorization state over time", and identity transitions
  • Consistent interpretation and correlation across domains and long-running workflows
  • Mechanisms to "collect, aggregate, verify, and transparently log records", possibly by a third-party auditor

The group is told to compose existing pieces: WIMSE for identity, RATS for attestation, OAuth for authorization, SCITT for transparency logging, W3C Trace Context for propagation, and vCon for conversation containers.

Post-hoc auditing comes first. Records can be created during an interaction, or before an action takes effect, but they are evaluated afterward. Real-time enforcement "may only be considered after re-charter."

What is out of scope?

Six things:

  • Auditing of model training
  • Real-time reasoning or enforcement
  • "How and when to create the records", and by which entity
  • Requirements on the auditing system itself
  • Session or context management, left to AGENTPROTO
  • "Authorization and delegation mechanism or extensions", left to OAuth. The charter text says "OATH working group". That is a typo to fix.

What are the deliverables?

  1. Architecture. An Informational RFC on roles, trust relationships, and data flows.
  2. Audit data models. "One or more" Standards Track RFCs that reuse or define models for interaction records, delegation context, authorization state over time, and action provenance. The vCon record format "should be considered". A registry for additional data models is on the table.
  3. Protocol extensions or profiles. "One or more" Standards Track RFCs for existing protocols, with examples of "HTTP headers, OAuth token formats, or attestation workflows".

The charter lists no milestones and no dates.

Where are the boundaries with other groups?

SCITT, RATS and vCon are clear. AUDIT profiles them. The BoF request says AUDIT would profile vCon's Verifiable Agent Conversation Records down to what third parties are allowed to see.

AGENTPROTO is clear on paper. Session management stays there. AUDIT wants to reuse a correlation identifier "to be specified by the agentproto WG". That is a dependency on something that does not exist yet.

OAuth is where the charter contradicts itself. The out-of-scope list excludes authorization and delegation extensions. The scope list includes attribution "by extending or adapting existing identity and authorization mechanisms". Deliverable 3 names OAuth token formats. The BoF request lists "context in tokens" as a possible OAuth change. The tiebreaker is one sentence: whether the work happens in AUDIT or the protocol's own group "needs to be considered on a case by case basis."

Here is a split that would work. OAuth keeps delegation, token issuance, token exchange, and the act claim. OAuth also keeps txn, which Transaction Tokens already defines as a transaction identifier. AUDIT gets the record that outlives the token, the store, and the retrieval path. Any new claim inside a token is an OAuth document. The charter should say that.

WIMSE is named as a building block. The BoF request lists it as a technology overlap. The BoF agenda's related-work line leaves it out.

Which reviewed drafts would land here?

These all define a signed record of an authorization or an action:

The first four straddle the line. They put evidence in OAuth tokens or issue it from the authorization server. The charter does not say which half of each belongs to AUDIT.

Is the charter tight enough?

No. Three things make it a catch-all.

It standardizes record formats while declining to say who writes them. Record creation is out of scope. The work is supposed to be "record-type agnostic". Deliverable 2 is a Standards Track data model with a "minimal set of audit information". A format with no required producer and no required moment of creation does not give an auditor a complete trail.

"One or more" appears twice, with no milestones. Any agent evidence draft fits under deliverable 2 or 3.

The one new protocol has no deliverable. The BoF request lists "Protocol(s) to store and retrieve records in the record store" as new work. The scope mentions collecting and logging records. The deliverables only cover profiles and extensions of existing protocols.

Two more problems. The charter sets the bar at verification by a party that trusts nobody in the chain. The architecture deliverable only says it "might" describe verification processes. Deliverables 2 and 3 both end with the same sentence about data representations for non-IETF logging systems. That reads as a paste error.

What does the BoF need to decide?

Whether the problem is specific enough for a working group, and whether this charter is the right starting text.

Questions worth asking at the mic:

  • Which group owns a new claim in an OAuth token that exists only for audit? Can the charter state a rule instead of "case by case"?
  • If record creation is out of scope, how does an auditor know a trail is complete?
  • Is the record store and retrieval protocol in the charter? Which deliverable covers it?
  • What is the first Standards Track document, and when is it due?
  • What does AUDIT use for correlation if AGENTPROTO does not specify an identifier in time? Is txn or W3C Trace Context the fallback?