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

Review · November 16, 2026

ARD Manifest and Lookup for Agent Discovery

draft-bu-dawn-ard-discovery -00View on Datatracker →

The Problem

An agent needs to find MCP servers, A2A agents and skills it was not configured with. This draft defines the catalog format a publisher serves and the HTTP search API a registry exposes.

How It Fits In

This is one of three Agent Resource Discovery (ARD) documents sharing a 10 minute slot on Thursday. The architecture names the roles. Initial contact finds a manifest URL from a domain name. This draft is everything after that.

It is a candidate for the charter's Protocol deliverable.

The manifest half fits the charter. The charter says an AI resource may be "a service such as a catalog that will further the discovery process", and that discovery "may use referrals, direct transfer, delegation, or indirection".

The lookup half does not fit.

  • The charter says DAWN "does not include open-ended search, semantic matchmaking, ranking, or selection of an AI resource based on capabilities or user intent". POST /search takes a natural language intent and returns scored results (Section 4.2.2).
  • "AI resource indexing and searching across the wider Internet" is out of scope. Section 4.1.1 describes registries crawling manifests across domains.
  • "Trust management and trust evaluation methods beyond secure content exchange" are out of scope. Third-party attestations and auditor signatures (Section 3.5.1) go past that line.

The charter also says work items "will provide integrity of all the discovery information exchanged". The trust manifest is optional, so an unsigned entry has transport security only.

Summary

An ARD entry describes one resource (Section 3.2). Four things are required:

  • identifier, a URN of the form urn:air:<publisher>:<namespace>:<agent-name>, where <publisher> is a domain name
  • displayName
  • type, a media type such as application/a2a-agent-card+json
  • exactly one of url (a link to the descriptor) or data (the descriptor inline)

The entry is an envelope. It points at an Agent Card or MCP server card and does not replace it. Optional fields include representativeQueries, capabilities, tags, version and digest. A manifest is a JSON document with an entries array (Section 3.6).

A trust manifest is an optional signed object inside an entry (Section 3.5). It binds a signer, such as did:web:example.com, to the entry's identifier, type, version and digest.

A registry exposes POST /search, POST /explore and GET /agents (Section 4.2). A federation parameter picks between forwarding the query to peers, returning referrals, or answering locally (Section 4.3).

What does it say about authentication and authorization?

Nothing about how to call the resource. There is no field for an authentication scheme, an authorization server, or scopes. The architecture says authentication and access control "are handled by the resource endpoint during invocation".

The trust manifest is provenance. It answers "did this domain publish this entry", not "may I call this".

The registry API has 401 and 403 errors (Section 4.2.5) and no defined way to authenticate.

How does it compare with MDI?

draft-cui-dawn-mdi-model is the abstract model for what the charter calls the minimum public discovery information. Its -02 requires a Name and an Endpoint.

  • ARD's url matches an MDI Endpoint whose purpose is description retrieval. An ARD entry never carries an endpoint for interacting with the resource.
  • ARD requires identifier and displayName. MDI makes both optional.
  • MDI has an Authentication Hint. ARD has none.
  • capabilities and the trust manifest are what MDI calls information beyond the minimum.

Neither draft cites the other.

Interoperability

Significant interoperability concerns

The ARD documents contradict each other on trust. Section 6.2 says fetching an entry from https://{domain}/.well-known/ard.json, "even when {domain} matches <publisher>", MUST NOT be treated as proof of publisher authority. The architecture (Section 4.2) says same-origin HTTPS "establishes domain authority" in exactly that case. One answer makes signatures mandatory for every publisher. The other makes a plain JSON file enough.

Verifying a trust manifest is not specified. The signer must match or be "cryptographically authorized by" the publisher, "e.g." did:web or X.509 SAN certificates. No method is mandatory. There is no reference for DIDs and no rule for resolving kid to a key.

The lookup API is a sketch. Section 4.3 requires a hopCount, and Section 6.6 requires a Hop-Count header and a "query trail header". None has a defined syntax. GET /agents has no response format. /explore has one example and no schema.

The prose and the CDDL disagree. Section 5 calls the CDDL normative. It makes subject and signature optional where Section 3.5 says MUST.

Mechanisms that could be removed

  • Recursive auto federation.
  • /explore and /agents.
  • Attestations, provenance links and third-party trust manifests.

Writing Quality

The examples are good. The text is not yet tight enough to implement from.

  • Section 4.3.1 says neither party "MUST rely on transport trust". It means MUST NOT.
  • The reference to initial contact uses the wrong draft name, draft-jimenez-ard-initial-contact.
  • application/ai-registry+json is registered as a media type with no document format defined for it.

Authors

From their Datatracker profiles:

  • Junjie Bu (Google). No RFCs. 3 active drafts, the three ARD documents.
  • R.V. Guha (no affiliation listed). No RFCs. 2 active drafts, both ARD.
  • Shaun Smith (Hugging Face). No RFCs. This is his only active draft.
  • Jaime Jimenez (Ericsson). 2 RFCs. 11 active drafts. CoRE chair.
  • Jeffrey Damick (Amazon). No RFCs. 2 active drafts.

Questions for the Working Group

  • The charter excludes semantic search, ranking and Internet-wide indexing, so should the authors split the manifest from the search API?
  • If the manifest is the candidate, should it be a binding of the MDI model?
  • Does a manifest served from the publisher's own domain need a signature?
  • Should a discovery record carry an authentication hint, as MDI allows, or leave that to the Agent Card and the resource?