ARD Specification: How AI Agents Discover Tools at Runtime

Agents know how to call tools. They know how to talk to other agents. What they have never had is a standard way to discover what tools and agents exist in the first place. That gap — the missing discovery layer — is what the Agentic Resource Discovery (ARD) specification addresses. On June 17, GitHub, Google, Microsoft, Hugging Face, GoDaddy, Cisco, Databricks, NVIDIA, Salesforce, ServiceNow, and Snowflake published the spec together. That coalition is not an accident.

What ARD Does

ARD sits above MCP and A2A in the agent stack. The division of responsibility is clean:

• ARD — answers “what tools and agents exist?”
• MCP — answers “how do I call a tool?”
• A2A — answers “how do I delegate to another agent?”

Without ARD, developers manually enumerate every capability an orchestrator can reach. That is fine for five tools. It breaks at five hundred. At enterprise scale — thousands of agents across dozens of teams — manual tool lists are not a process, they are a liability.

The specification operates in four steps. A resource publisher creates an ai-catalog.json manifest at a well-known URL on their domain, describing what the tool does and how to invoke it. A discovery service indexes that manifest. An orchestrator queries the discovery service in natural language — “find me a tool that processes invoices” — and gets back ranked results. The orchestrator then connects to the selected resource over its native protocol: MCP, A2A, or plain REST.

The spec is protocol-agnostic by design. It does not replace MCP or A2A. It tells agents where to look before those protocols take over.

The Coalition Is the Real Signal

The technical design of ARD is reasonable. What is extraordinary is who signed it.

GitHub, Google, Microsoft, Hugging Face, and GoDaddy co-authored the spec. Cisco, Databricks, NVIDIA, Salesforce, ServiceNow, and Snowflake came in as supporting partners. Every major AI platform is at the table. When something like this last happened in developer infrastructure — think OAuth 2.0 in 2011 — the standard won. Not always because it was the best design, but because the whole ecosystem moved together.

ARD is currently at version 0.9. It will change before 1.0. But the coalition signals that whatever shape 1.0 takes, this is the spec developers will build against. Google’s announcement post names the full partner list if you want to verify the depth of that commitment.

Two Reference Implementations, Live Today

Announcing a spec and shipping working code are different things. ARD shipped both on day one.

GitHub Agent Finder is now available on all GitHub Copilot plans. Instead of pre-loading every possible tool into an agent’s context, Copilot queries GitHub’s catalog at runtime to find the right MCP servers, skills, and agents for the current task. The practical benefit is significant: a 10,000-tool catalog pre-loaded into context is not functional. Injecting only what is relevant is.

Hugging Face hf-discover is the open-source reference implementation. The CLI is straightforward:

hf discover 'help me train a model' --json

It searches Hugging Face’s catalog of Skills, ML applications, and MCP servers, and returns results as typed artifacts. The registry is federated — it searches across Hugging Face and other ARD-compliant services simultaneously. Hugging Face’s launch post covers the full implementation details.

Snowflake announced that Cortex Agents will automatically register into an organization’s ARD discovery endpoint. Cisco published AGNTCY Agent Directory, an open-source enterprise implementation for cross-boundary discovery with local governance controls.

What Developers Should Do Now

ARD is a draft spec at v0.9. Early adoption carries real risk — the manifest format could shift before 1.0. With that caveat stated plainly, here is the practical action list:

• If you publish tools or agents for others to use, publish an ai-catalog.json now. Getting into catalogs early means being found when discovery becomes table stakes.
• If you are on any GitHub Copilot plan, Agent Finder is already available. Test it against your workflow.
• Run hf discover to understand what the search experience looks like from a consumer’s perspective.
• Watch the ards-project/ard-spec repository for v1.0 changes before committing to production integrations.

The DNS analogy holds here. The web did not collapse because developers had to maintain IP address lists — but it scaled because DNS made that unnecessary. ARD is betting that agent infrastructure follows the same arc. Given who is behind it, that is not an unreasonable bet.