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

Review · November 21, 2026

Verifiable Mandates for Agents Acting Across Administrative Domains

draft-hulea-wimse-agent-mandate -01View on Datatracker →

The Problem

A company's treasury agent asks the company's bank to make a payment. The bank has to decide whether to accept an action from an agent that was configured and authorized somewhere else.

The delegation chain drafts already cover part of this. They prove that each hop in a chain narrowed the permissions it received. They don't prove that the person at the top of the chain was allowed to grant those permissions in the first place.

This draft adds that missing piece. A third party, such as a company register, signs a statement of which powers a person may delegate to agents. The bank checks the mandate against that statement.

How It Fits In

This is a profile of Token Exchange that uses Rich Authorization Requests objects as its permission language. It has 14 normative RFC references plus three working group drafts: WIMSE Workload Credentials, WIMSE HTTP Signatures and Token Status List. It also makes JSON Schema 2020-12 a normative dependency.

It adds roles that don't exist in OAuth or WIMSE today: Authority Source, Mandate Issuer, Delegator, Validator, an "observer" that watches agent sessions, and a Trust Framework that sits above all of them.

Does it belong in WIMSE? Only partly. The agent authenticates with a WIMSE Workload Identity Token and signs requests with the WIMSE HTTP Signatures profile. Everything else is OAuth: a new token type, a new token endpoint parameter, RAR matching rules and access token requirements. The statements about legal representatives and eIDAS attestations of powers are closer to digital credentials work than to workload identity.

It doesn't mention Identity and Authorization Chaining Across Domains, the adopted OAuth work on cross-domain token exchange. That's the first comparison the OAuth working group will ask for.

The draft overlaps heavily with Verifiable Attenuated Delegation for AI Agent Chains, and the author says so. The stated plan is to drop this draft's own chain format in the next revision and carry the chain in that format instead.

What It Defines

Authority Assertion. A JWT from an Authority Source. It says a named delegator may delegate a set of powers to agents on behalf of a named principal. It says nothing about whether any mandate was ever granted.

Mandate Assertion. A JWT from a Mandate Issuer, bound to the agent's key. It names the principal and delegator, lists permissions as RAR-shaped objects, and references the Authority Assertion by hash. An agent can sign a narrower mandate for another agent, up to max_depth.

Constraints. Each permission can carry constraints that point into the request with a JSON Pointer. There are seven predicates: eq, in, max, min, window, sum_max and cosign_above. The last two are stateful. sum_max is a cumulative budget over a period. cosign_above requires a second approver above a threshold.

Presentation. The agent makes a token exchange request to the relying party's authorization server. The leaf mandate is the subject_token. The WIT is the actor_token. A new mandate_evidence parameter carries the rest of the chain and the Authority Assertions. authorization_details carries exactly one requested action. The whole request is signed with HTTP Message Signatures.

Validation. A Validator inside the authorization server runs a 13-step procedure. Its inputs are fixed: the presentation, a Validation Profile, a Trust Anchor Set, status list tokens and the time. It outputs a Validation Result of valid, invalid or indeterminate, with every claim attributed to the artifact and role that asserted it. Two validators given the same inputs must produce the same result.

Authorization. The relying party applies its own policy to the result. It has to discharge the stateful constraints itself, for example by atomically recording the payment against the budget. It then issues a sender-constrained access token good for one execution of one action.

Optional evidence. The agent can add a RATS attestation result. It can also add an Exposure Statement, a new single-use JWT from the observer listing whether the session touched external content, unauthenticated tools or unvalidated agents.

Significant Interoperability Concerns

Everything depends on a trust framework that doesn't exist. The framework has to define twelve things, including the accepted issuers, the identifier schemes, the RAR types and how budgets are shared. The draft is a toolkit for writing such a framework. Two implementations of this draft alone can't interoperate.

The Authority Source has to speak the relying party's API vocabulary. powers are RAR objects with JSON Pointers into the request. In the example, a company register signs a statement about /instructedAmount/amount on a payment_initiation type. Registers don't record powers this way today. They also don't record which powers are delegable to agents, which the draft treats as a separate fact from what the person may do themselves.

The link between the delegator and the mandate isn't verified. The Authority Assertion says the director may delegate. The Mandate Issuer says the director did. Nothing in the protocol ties the two together. The draft admits this and leaves it to the trust framework. The relying party already has to trust the Mandate Issuer as an anchor, so the Authority Assertion only caps what that issuer can claim.

The access token is underspecified. It must be sender-constrained, but the draft doesn't say whether that means DPoP, mTLS or WIMSE signatures. It must carry the jti of the leaf and root mandates, but no claim names are defined. Single-use enforcement is left to the resource server with no mechanism.

The agent's status as an OAuth client isn't defined. The draft says the agent instance is the client. It doesn't define a client authentication method or say what the client_id is.

Budgets are weaker than they look. A budget is counted per relying party and per root mandate. A mandate accepted at two banks gets the full budget at each. A reissued root mandate starts a fresh budget. The draft says both things plainly, and recommends a single audience for mandates with budgets.

The agent chooses which evidence to present. A root mandate can reference several Authority Assertions as alternatives. If one is revoked, the agent leaves it out and presents another. The relying party sees an authority_not_presented diagnostic and nothing else.

Type Profiles live in the relying party's Validation Profile. The mandate is written earlier by someone else. The Mandate Issuer has to know which members the relying party's profile admits, or the mandate won't match any request. The profile is "usually" issued by the relying party, so each bank can admit different members for the same type.

Mechanisms That Could Be Removed

  • The chain format and sub-delegation. The author already plans to move this to the Asor draft. Doing it now would cut a large part of the document.
  • Exposure Statements. A component run by the agent's own operator reports on the agent's session. The consumption rules take a page. The draft itself says an empty list doesn't mean the session was clean. This is a separate draft at most.
  • The SD-JWT variant of the Authority Assertion. Nothing that affects validation may be selectively disclosed, and disclosures are ignored. A plain JWT does the same job.
  • Attestation results. The relying party can require an EAT by policy without this draft profiling it.
  • Byte-identical Validation Results. A lot of text exists to make two validators produce the same JSON: sort orders, tie-breaking between status tokens, number comparison rules, 37 registered codes. The Validator runs inside the relying party's own authorization server. The decision record that would let a third party use the result is listed as an open issue. This machinery is ahead of its use case.
  • JSON Schema as a normative dependency. The relying party can validate its own request format however it likes.

The core idea survives all of these cuts. It is one new JWT type and one check: the root mandate's permissions fit within the Authority Assertion's powers.

Writing Quality

The writing is good. Sentences are short and precise. Normative language is used correctly. The non-goals and the trust framework boundary are stated clearly, and the security considerations are candid about what the design doesn't solve.

The problem is size. This is 57 pages with four JWT types, three new IANA registries, eleven new JWT claims and nearly 50 conformance cases. The -01 revision came nine days after -00 and its change list runs more than a page. The validation procedure is detailed down to which error code wins when a status token fails two checks.

That level of detail belongs at the end of a working group process. At this stage it makes the one new idea hard to find and hard to review.

Appendix C is useful. It maps this draft against nine related drafts and gives a predicate-by-predicate comparison with the Asor constraint types. The acknowledgments show the author is working with the other individual draft authors in this space.

Background of the Author

Claudiu Hulea (Braincap SRL, Bucharest) is a new IETF author. He has no RFCs and no expired drafts. This is his only active Internet-Draft.