Agentic Resource Discovery
What is ARD?
ARD stands for Agentic Resource Discovery. It is an open specification for how AI agents find the tools, skills, agents and APIs they need, published in June 2026 by a working group including Google, Microsoft, Hugging Face, AWS, Cisco, GitHub, Nvidia, Salesforce and Snowflake. It answers one question, “what is available for this task?”, and then gets out of the way.
The problem ARD solves
Today an agent can only use capabilities that somebody wired in beforehand. Every tool has to be integrated by hand, and every tool description has to sit in the context window competing for space with the actual work. That model does not survive an ecosystem of thousands of tools, let alone millions.
ARD moves the selection problem out of the context window and into a search service. It is the same shift the early web made when it went from curated link directories to search engines: you stop maintaining a list of everything and start asking a question.
How ARD works
There are two sides, and the specification is deliberately small on both.
Publishers describe what they offer
A publisher serves a JSON manifest at /.well-known/ard.json on their own domain. Each
entry names one resource: an identifier, a display name, an IANA media type saying what kind of thing
it is, a URL, a description, and representative queries phrased the way a person would ask.
Nobody approves this. There is no marketplace to apply to and no allowlist. Publishing on your own domain is the whole mechanism, which is why the specification calls the model web-native.
Registries index and answer
A registry crawls those manifests and exposes a search API. An agent asks in plain language, the registry returns ranked candidates, and the agent connects to whichever it picks.
POST /search
{"query": {"text": "scrape a website behind cloudflare"}}
The four discovery paths
A publisher can advertise entries four ways, and a consumer may check any of them. Serving only one makes you invisible to any client that checks another.
| Path | What it is |
|---|---|
/.well-known/ard.json | The manifest itself. A conformant consumer MUST fetch this one. |
Agentmap: in robots.txt | The agent-facing counterpart to Sitemap:. |
<link rel="ard"> | A tag in the page head, for crawlers that only read HTML. |
| DNS service records | Points at a static entry source or a live registry endpoint. |
The predecessor path /.well-known/ai-catalog.json is still read by
crawlers written against the earlier draft. Serving both costs one route.
Federation
The specification defines three federation modes and makes auto the default: a
registry queries its peers, merges their results with its own, and returns one set. The other two are
referrals, where a registry returns pointers to its peers, and none, where
it answers only from its own index.
In practice, when we benchmarked every public ARD registry on 31 August 2026, none of them
implemented the default mode. Asking for auto returned byte-identical results to
none on all four.
Trust is deliberately separate
ARD is explicit that a relevance score reflects semantic relevance only and must not be read as a trust, compliance or safety judgment. Trust evaluation is decoupled and handled through a separate trust manifest, so a registry's ranking can never be mistaken for an endorsement.
Check your own domain. The console fetches what you publish and asks every public registry whether they return you.
Run a free audit