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

Review · November 21, 2026

Delegator-Bound Delegation Chains for AI Agents

draft-nandakumar-wimse-delegator-bound-chains -00View on Datatracker →

The Problem

draft-asor-wimse-agent-delegation-chain lets a resource server check offline that each token in a chain of agent delegations carries no more authority than its parent. It never says which key signs a child token.

So a valid chain proves each hop is narrower than the last. It does not prove the holder of the parent token granted the child.

This draft fixes that. Each child token must be signed with the key its parent confirms in cnf. It then adds four more rules on top: a pseudonymous principal, cycle detection, audience narrowing, and a context claim.

How It Fits In

This is an extension to another individual draft, written by different authors. Its only normative draft reference is the Asor draft. The other eight normative references are JOSE, JWT, and HMAC RFCs.

It cites no WIMSE working group document. The base draft profiles JWT access tokens, RAR, and DPoP, so this is OAuth token design submitted to WIMSE. Agent-to-agent delegation without an authorization server is not in the WIMSE program of work.

It does change the roles. In OAuth, only an authorization server issues access tokens. Here every agent holding a token becomes an issuer of at+jwt tokens, signed with its DPoP key. The authorization server appears once, at the root.

The draft positions itself as a "companion profile" to draft-mcguinness-oauth-actor-profile, supplying per-hop evidence that the nested act claim doesn't carry. It also maps its rules onto Delegate SD-JWT, which the OAuth working group has adopted and which already signs each hop with the previous hop's key.

How It Works

The root token DT_0 comes from an authorization server. It names agent-a in client_id and agent-a's key in cnf. If it carries del_bind: delegator, the rules below apply to the whole chain.

To delegate, agent-a mints DT_1 itself:

  • It signs DT_1 with the key from DT_0's cnf
  • It sets iss to its own client_id
  • It names agent-b in client_id and agent-b's key in cnf
  • It narrows authorization_details and aud

Agent-b does the same for agent-c. The resource server gets the full chain, verifies DT_0 against a trusted root key, then verifies each child's signature with the key from its parent. The last agent proves possession of the last key.

The four additions:

  • Originator pseudonym. sub can be an HMAC of the principal's identifier and the root token's jti. A new sub_kind claim marks it. Only the root issuer can map it back to a person.
  • Cycle detection. A verifier rejects a chain where the same client_id appears twice.
  • Audience attenuation. A child's aud must be a subset of its parent's.
  • Context binding. A ctx claim names a dialog or a Transaction Token txn. A receiver that knows its context rejects a mismatch in the last token.

Significant Interoperability Concerns

iss now means two different things. The base draft defines iss "as in RFC 9068", where it is the authorization server's issuer identifier and the verifier uses it to find keys. This draft sets it to a client_id string in every child. A stock JWT access token validator will reject these tokens. The draft doesn't acknowledge the conflict.

Agent names are unauthenticated. The signature proves that the holder of key A issued DT_1. The string "agent-b" in DT_1 is whatever agent-a typed. The iss equals client_id check compares two strings chosen by the same party. The chain proves a chain of keys, not a chain of identified agents. The claim that this "attributes" tokens to an acting agent only holds for the first hop.

Cycle detection relies on those same strings. client_id values have no shared namespace across organizations. A delegator can rename the next agent and the check passes. The draft admits per-chain identifiers defeat it. The rule also blocks legitimate patterns, such as agent-b calling back into agent-a. This catches bugs among cooperating agents. It is not a security property.

The pseudonym breaks the resource server's job. In the example, a payment service receives a per-chain pseudonym for the person. It cannot map that to an account. The draft doesn't say how a resource server acts for a principal it can't identify.

Audience narrowing requires the root to know every destination. DT_0 must list every resource any downstream agent will ever call. The draft targets cross-organization delegation where agents pick sub-agents dynamically. Those two goals conflict.

ctx binds less than it claims. Any delegator can add or change it, and only the last token's value is checked. It is the last delegator's assertion, not a property of the chain. The draft also doesn't say what happens when a receiver knows its context and ctx is absent.

Revocation has no owner. The security considerations lean on Token Status List. Child tokens are issued by agents. Nothing says who hosts a status list for a token that an agent minted with its DPoP key.

The base draft's gaps are inherited. There is still no way to present a chain in an HTTP request, and no defined way for agent-a to learn agent-b's key. The base draft's Datatracker history also lists an IPR disclosure.

One key does two jobs. The DPoP key now signs access tokens too. The draft's answer for deployments that want separate keys is "do not use this extension".

Mechanisms That Could Be Removed

  • The HMAC formula. Only the root issuer computes the pseudonym, and nobody else can verify it. A random value and a lookup table interoperate identically. HMAC also isn't reversible, so "resolving" it needs a stored mapping anyway.
  • sub_kind. No verifier behavior depends on its value. It is only checked for sameness.
  • del_bind in every child. DT_0 is signed by the root and pinned by par_hash. A delegator can't remove it. Repeating it adds a rejection rule and nothing else.
  • The Delegate SD-JWT section. It is half a page with no normative keywords, profiling an OAuth working group document. delegate_payload is an array there, and the mapping doesn't say which element carries what.
  • ctx, until it has a rule that covers more than the last hop.

Writing Quality

The writing is good. It is 12 pages of content, with correct IETF conventions and short, clear sentences. The worked example in Section 2 comes before the rules and makes the design easy to follow. Section 3.4 states plainly what a verified chain proves.

There are small slips. The example says ctx "names the dialog each token was issued for", but the figure shows no ctx. The draft mentions "an earlier proposal by the authors" without naming it. Three terse JWT claim names (ctx, sub_kind, del_bind) go into a global registry on the strength of an individual draft.

The bigger issue is structural. Delegator binding is a fix to a hole in the base draft. The other four rules are separate features that happen to ride along. Each needs its own justification and gets about half a page.

Authors

Suhas Nandakumar (Cisco) has 9 published RFCs from 2013 to 2025, all in real-time media: STUN and TURN URI schemes, RTP taxonomy, SDP and ICE. He has 36 active Internet-Drafts. Most are Media over QUIC. About ten are recent OAuth, WIMSE, and AGENTPROTO drafts. He is an experienced IETF author who is new to this area and writing at high volume right now.

Cullen Fluffy Jennings (Cisco) has 29 published RFCs from 2002 to 2026, including SIP identity (RFC 3325, RFC 4474) and NAT behavior (RFC 4787). He has 30 active Internet-Drafts, mostly Media over QUIC plus the same set of agent, OAuth, and WIMSE drafts co-authored with Nandakumar. He is a long-time IETF contributor, also new to these working groups.

This draft is one of a set. The same authors' delegation revocation draft depends on its key resolution rules.

Bottom Line

The core rule is right. A child token signed with the key its parent confirms is the answer the base draft was missing, and Delegate SD-JWT already works this way.

So where should it live? In the base draft, not in an optional extension. A chain format where "who signs the child" depends on an opt-in claim from a second document is the wrong shape. The best outcome is that draft-asor-wimse-agent-delegation-chain absorbs Section 3 and resolves the iss conflict with RFC 9068.

The other four mechanisms need more work before they are worth standardizing. Audience narrowing is the strongest of them. The pseudonym, cycle detection, and ctx all depend on strings that no one in the chain can verify.