IETF draft maps agent approvals and refusals onto signed audit receipts

A new revision of an Internet-Draft published on 15 November turns the IETF’s emerging architecture for auditing agent delegation into concrete, signed records. The draft maps two receipt types — a decision receipt and an outcome receipt — onto the architecture’s Interaction, Action and Authorization Transition Records, for three cases: a human approval, a declined action, and an executed action with its result.

What happened

K. Gruszka published revision -08 of draft-gruszka-audit-decision-receipts, “Decision Receipts for Approvals, Declined Actions and Executed Actions in Agent Audit Records”, as an individual Informational Internet-Draft. It is 60 pages, including test vectors, and it is a submission to the IETF’s AUDIT effort rather than an adopted working-group document.

The context is draft-kuehlewind-audit-architecture, “An Architecture for Auditing Agent Delegation and Interactions”, which proposes that an audit link user intent, delegation, authorization and execution, and treats authorization as a state that changes over time. That architecture names four record types and leaves their encoding open; a canonical data model is an explicit work item. The new draft fills that gap for three events using a receipt format that already exists.

A decision receipt is an in-toto Statement inside a DSSE envelope, signed with Ed25519 over JSON canonicalised under RFC 8785. The draft binds every input by SHA-256 digest: the action bytes, the approval record, the policy boundary, the evidence, the decision itself and the observed effect. A decision receipt is created when the decision is taken, before the action runs. An outcome receipt is created afterwards by the executor or by a separate observer.

The three cases are treated as records in their own right. A human approval becomes an Interaction Record plus an Authorization Transition Record. A declined action becomes a decision receipt with a DENY or REFUSE verdict, with its reason and without an effect. An executed action becomes an Action Record, with the outcome receipt naming the authorising decision by digest. A refusal is not an absence of logs; it is a record.

The draft also specifies an offline verification procedure that a third party can run over the exact bytes, and it spends a full section on one control that most approval gates get wrong: an approval may be honoured at most once. Consumption state is keyed by the identity of the approval — approvalId, actionDigest and session — not by the bytes of one copy. A copy with a later validUntil is the same approval, and extending validity needs a new approval act and a new approvalId. The consumption mark must be durable before the action runs or before a signed ALLOW leaves the relying party’s control, and it must survive rotation, archiving and compaction. If the state cannot be read, the relying party must refuse.

Implementation status is documented per RFC 7942. The proofbundle project, maintained by the draft’s author, is on PyPI at version 6.4.0 under the MIT licence and produces and verifies the receipts in the test vectors; the draft notes that 6.5.0 was the latest release on 15 November. Agent Guard, a preview plugin for Claude Code and Codex, provides the gate; a Go verifier and registration with a local scitt-ccf-ledger are also described. Registration with a Transparency Service under RFC 9943 is covered informatively, including what a SCITT Receipt would and would not show.

The limits are stated plainly. A receipt does not show that a decision was correct, safe or lawful. It does not authenticate the approval: the approval record in the draft is not signed by the person it names, and the draft points to separately signed approvals and user-verification-gated tokens as possible additional evidence without defining a mapping. It does not prove an effect happened, only that a signer stated an effect with a given digest. It is not an access token or an authorization grant. Its times are the signer’s clock, and it cannot detect a receipt that was never written.

Why it matters for KYA

Most agent audit trails today are log lines written by the agent’s own operator. That is a record of what the operator says happened. An auditor cannot distinguish “a human approved this action” from “the agent wrote that a human approved it” if the record is unsigned, mutable, and produced by the same system that took the action. This draft is useful because it treats that distinction as the design problem rather than an implementation detail.

The separation of decision from outcome is the part worth copying even if you never implement this format. Authorisation is a decision taken before an action, with a named approver, a scope and a validity window. The effect is a separate claim, signed by a different party, binding the result to the decision by digest. Those are the two questions a KYA review asks after an incident: was this action authorised, and did it actually happen. A single log line answers neither.

Single-use approvals are the control enterprises most often miss. An approval that can be replayed, or a copy whose validity can be extended after the fact, is a delegation that outlives the intent behind it. The draft’s insistence that consumption survive rotation is the unglamorous engineering that decides whether an approval gate is real. Recording declines matters for the same reason: if only executed actions are logged, nobody can show that a risky action was refused.

The honest limits are also a useful checklist for buyers. If a vendor claims “cryptographically verifiable agent audit”, ask who signs the approval, whether the approver’s own signature is bound, whether the effect is attested by an independent party, and whether the record can prove that a receipt was not withheld. This draft is explicit that its receipts do not answer all four.

What to watch

  • Whether the AUDIT BoF, meeting at IETF 127 this week, gets a working group and adopts a canonical data model. The encoding question is where interoperability will be won or lost.
  • Whether a signed approval record follows. The draft’s weakest link is that the human act is bound by digest but not signed by the human.
  • Whether transparency-service registration becomes routine for high-stakes actions, or stays an optional extra that only regulated deployments configure.
  • Whether agent platforms adopt receipt-shaped logs, or keep shipping free-text traces that cannot be checked by anyone outside the vendor.

Sources

  • K. Gruszka, “Decision Receipts for Approvals, Declined Actions and Executed Actions in Agent Audit Records”, draft-gruszka-audit-decision-receipts-08, 15 November 2026 — datatracker.ietf.org
  • M. Kuehlewind and H. Birkholz, “An Architecture for Auditing Agent Delegation and Interactions”, draft-kuehlewind-audit-architecture-01, 7 September 2026 — datatracker.ietf.org
  • IETF, AUDIT BoF agenda, IETF 127 — datatracker.ietf.org
  • proofbundle, decision and outcome receipt predicates, v6.4.0 — github.com

Know Your Agent (KYA) explains agent identity, verification and accountability. This is an explainer, not legal or compliance advice — see our Sources & methodology. New to KYA? Start with What is Know Your Agent? and the glossary.