Review · November 16, 2026
Agent Resource Discovery (ARD) Architecture
The Problem
An agent cannot load every MCP server, A2A agent and skill into its context window. The tool list has to be found at runtime, across organizations, without one company running the directory.
ARD's answer is a search layer. Publishers describe their resources in a JSON manifest. Registries index those manifests. A consumer sends a query and gets back the top few entries.
This draft is the overview for a three-document set. It has no normative requirements and no IANA actions.
How It Fits In
ARD shares one 10 minute slot on Thursday across three drafts. This one covers roles and the trust model. The companions are:
- Initial Contact: how a client gets from a domain name or a local network to a manifest URI.
/.well-known/ard.jsonis mandatory. DNS and mDNS pointers are optional. - Manifest and Lookup: the
application/ard+jsonentry format, the signedtrustManifest, and the registry search API with federation.
Registration is "to be specified in a later document" (Section 1).
So does it fit the charter? Partly.
The three deployment settings in Section 3.2 line up with the charter's three contexts. The charter allows the discovered resource to be "a service such as a catalog that will further the discovery process". A manifest that points at a registry is that.
Search 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". ARD's second principle is "Search-First Dynamic Discovery" (Section 3.3.2).
Summary
There are three roles (Section 2):
- A publisher controls a domain and serves a manifest at
/.well-known/ard.json. - An ARD Registry holds entries. It fills itself by crawling manifests or by accepting registrations.
- A consumer reads a manifest directly or queries a registry, then follows the entry's
urlto the real descriptor, such as an A2A Agent Card or MCP server card.
The registry never sits in the runtime path (Section 3.3.7). Identifiers are URNs, urn:air:<publisher>:<namespace>:<agent-name>. The manifest and the agent can both be hosted by someone else.
That gives the three-domain trust model in Section 4.1. TLS authenticates whoever served the manifest or registry response. An optional trustManifest signature, with an identity like did:web:abc.com, ties the entry to the publisher. A digest pins the descriptor hosted on a third domain.
How it compares to the DNS proposals
DNS-AID puts each agent in an SVCB record and declares capability search out of scope. Its _index._agents.<domain> record points at a catalog whose format it does not define. One of its examples uses well-known=ard.json. ARD Initial Contact reuses the same owner name. So ARD and DNS-AID fit together.
DAN is the real contrast. It defines two new record types, puts endpoint, capabilities and certificate data in DNS, and requires DNSSEC validation. In DAN, trust comes from the signed zone. In ARD, DNS answers are "untrusted hints" (Section 4.2) and trust comes from TLS plus a per-entry signature.
Interoperability
Significant interoperability concerns
The three drafts disagree on when a publisher is verified. This draft says that an entry fetched from https://<publisher>/.well-known/ard.json is covered by same-origin HTTPS (Sections 4.1 and 4.2). The lookup draft's Section 6.2 says fetching from that URL, even when the domain matches the publisher, MUST NOT be treated as proof. It requires a verified trustManifest or the entry is an "unverified discovery hint". Those are opposite answers for the simplest deployment.
The signature that carries the trust model is optional. Section 3.1 says the publisher of an entry "is established from its trustManifest", and the same section makes it optional. The charter says work items "will provide integrity of all the discovery information exchanged". An optional signature does not meet that once a registry is in the middle.
The companions do not line up. This draft says registration is unspecified. The lookup draft says it is specified in [AGENT-DIRECTORY]. Initial Contact defines _index._agents.<domain>. The lookup draft's Section 4.1 describes it as _entries._agents and _search._agents.
Mechanisms that could be removed
- JSON-LD (Section 3.3.5). A registry of field names would do the same job.
- Third-party auditor signatures and attestations. The charter excludes trust evaluation beyond secure content exchange.
Writing Quality
This is the cleanest of the three ARD drafts. It is 15 pages, the figures are useful, and the definitions are tight.
It does not cite the DAWN terminology draft. Section 3.3.4 calls application/mcp-server-card+json and application/ai-skill+md "IANA Media Types", and no draft in the set registers them.
Authors
From their Datatracker profiles:
- Junjie Bu (Google). No RFCs. 3 active drafts, all ARD.
- R.V. Guha. No RFCs. 2 active drafts, both ARD.
- Jari Arkko (Ericsson). 58 RFCs. 3 active drafts, all in DAWN. Chair of PROCON.
- Jaime Jimenez (Ericsson). 2 RFCs. 11 active drafts across CoRE, T2TRG and DAWN. Chair of CoRE.
Questions for the Working Group
The session ends with a review of candidates against the deliverables. Here is where ARD lands:
- Terminology and Use Cases and Applicability: not addressed.
- Discovery Architecture: this draft is a candidate. It has no requirements and does not say what the minimum public discovery information is.
- Protocol: Initial Contact and the manifest fit the deliverable. The search API and cross-domain federation go past it.
So what should the WG decide?
- Does the WG take the manifest and Initial Contact and leave the search API for a later charter?
- Which fields of an ARD entry are the minimum public discovery information?
- Can ARD be the catalog behind DNS-AID's
_index._agentsrecord, so the WG adopts a DNS layer and an HTTP catalog layer together? - For a manifest served from the publisher's own domain, is same-origin HTTPS enough, or is a signature always required?