Standards Research

ARD — The Discovery Layer Its Own Authors Have Not Published

Eleven of the largest agent builders specified how the web should advertise its agentic resources. Nine companies did it, seven of them are outsiders, and none of them is fully conformant.

On the standard: Agentic Resource Discovery · Updated July 2026 · Evidence as of 31 July 2026

This is my standing research position on Agentic Resource Discovery — updated as I re-probe the catalog, and written to be argued with. The measured population lives free at apis.io/ard/, one page per publisher with the verbatim manifest each grade was computed from, and the working layer underneath both is The ARD Standard.

I am invested in this one rather than scoring it once and filing it, because ARD is aimed at the bottleneck I keep running into in every other report.

Why this standard matters more than its adoption suggests

Every agentic protocol shipped its invocation layer and skipped discovery. MCP wires an agent to tools. A2A lets agents delegate to each other. OpenAPI describes the HTTP underneath all of it. Not one of them tells a client where to look in the first place, which is why an agent’s world is still whatever a developer wired into it by hand.

ARD’s answer is the cheapest correct shape for that problem: a static JSON file at /.well-known/ai-catalog.json on your own domain, an identifier anchored to that domain, one mandatory POST /search endpoint, and trust kept deliberately separate from relevance. It inverts submit-to-a-registry into publish-to-your-own-domain, and it refuses to be the one registry. Several things about the draft are better than most standards manage in their first year:

• Three formal schemas with an explicit statement of which is authoritative — CDDL for the grammar, JSON Schema for runtime validation, OpenAPI for the interface.
• An official conformance CLI in the same repository as the specification, wired into CI. Every verdict in my research came from it.
• A public ADR ledger — nine records, one properly marked superseded, each with context, decision and consequences. Reading them is the fastest way to understand why the specification looks the way it does.
• score MUST NOT be read as trust. The specification says this in bold, and it is the most disciplined line in the document.

What I found probing it

Nine weeks after the v0.9 draft landed, I fetched /.well-known/ai-catalog.json on every absolute host in the API Evangelist catalog plus each host’s derived apex — 43,834 domains, 37,751 reachable — and graded every manifest twice: against the AI Catalog 1.0 data model, and with ARD’s own conformance tool run against the same bytes.

Nine publishers. Zapier, ClickHouse, Airtable, DataRobot, Roboflow, Hugging Face, Opus, Artie and Fundraise Up. None is fully conformant. Four pass the official tool.

The ratio runs backwards. Eleven organizations are on ARD’s contributors wall. Probed across 127 of their domains, two serve a manifest anywhere — Hugging Face, and Cisco, whose reference implementation fails ARD’s own conformance test. Seven of the nine publishers are not in the working group at all. Adoption is running ahead of authorship, which is the opposite of every other standard I have measured.

Two of the five conformance failures are the documentation’s fault, not the implementer’s. DataRobot and Opus emit urn:ai: identifiers — the scheme ARD itself specified until ADR-0009 renamed the URN namespace to urn:air: for validity. There is no tagged release anywhere in the organization, and four artifacts disagree about the version number, so there was nothing for them to watch. And Airtable fails on a root collections array the docs site still documents as a field.

Three of four discovery mechanisms are effectively unused. One robots.txt Agentmap directive (Roboflow), one HTML <link rel="ai-catalog"> (ClickHouse), and zero DNS records at any of the four labels the specification and the publishing guide disagree about. A methodology note, because this is where a survey goes wrong: wildcard DNS makes dig +short SRV answer on any domain, so the answer section has to be parsed by record type — one domain looked like a publisher until a made-up label returned the identical record.

Publishing a catalog does not correlate with being a better API provider. The nine publishers average 57.5 on the Kin Score and 44.0 on Agent Readiness; the eleven coalition companies average 57.0 and 45.1. Statistically the same population. Today, publishing an ai-catalog.json is a curiosity signal rather than a maturity signal — and that is exactly what I would expect nine weeks in.

Federation is specified, implemented and empty. Across the four reference registries, every referral anyone returns points at a second index run by the same operator. The Public Agent Finder in every federation example in the specification, finder.nlweb.ai, does not resolve.

The best manifest in the whole corpus belongs to Roboflow — a did:web: host identity, a real SOC 2 attestation, an honest signatureTodo note about pending key material, and two discovery mechanisms — and Roboflow was not in my catalog at all until this probe put it there.

Where I think this goes

The specification is not the problem. The loop between publishing and being read is the problem. Nine companies published a catalog into a web where four registries exist, one of them cannot be tested by the official tool because its edge blocks non-browser clients, none refers a client to another vendor, and no client queries any of them unless a developer wires it up by hand. That is the same failure mode Schema.org avoided by making consumption visible and valuable — and one of ARD’s authors is the person who built Schema.org, so the lesson is available in-house.

Three unglamorous fixes would change my read: tag a version, correct the three pages where the documentation contradicts the tooling, and publish who passes. All three are days of work, not quarters.

What I am building against the gaps

• A conformance register. Nobody publishes ARD conformance results anywhere, in any machine-readable form. Every grade behind apis.io/ard/ is dated, carries the verbatim manifest it was computed from, and is regenerated on every build — which makes it the beginning of one.
• Manifest generation rather than authoring. Four of the nine publishers carry no representativeQueries and two have no descriptions, because every one of these files was written by hand. An OpenAPI, an MCP server list and an AGENTS.md contain nearly everything a good ai-catalog.json needs. This is a generator, and it is the reason most of the corpus is near-conformant rather than conformant.
• A drift monitor. Two publishers are stranded on a superseded identifier scheme because there was no version to subscribe to. Diffing a published manifest against the current draft and telling the publisher what moved is a service; this research is a manual first run of it.
• A scoring decision I have not made yet. The AI Catalog manifest is, like the Agent Card, an artifact that cannot be derived on a provider’s behalf — which is exactly the property that earns a dimension in the Agent Readiness rubric. With nine publishers in 37,751 domains, adding it today would rescale the whole catalog to measure something almost nobody does. I am holding it until the population justifies it, and I would rather say so than quietly weight it.

Corrections

Everything here is dated, and every number came from a public artifact I can point at. This is a v0.9 draft moving quickly. If you work on ARD, publish a catalog, or run one of the registries and I have read your surface wrong, pointing me at the artifact is the fastest way to change this page — it exists to be corrected. [email protected].