Agentic Resource Discovery: Let agents search

Source: HuggingFace Date: 2026-06-17 URL: https://huggingface.co/blog/agentic-resource-discovery-launch

Summary

Agentic Resource Discovery (ARD) is an open specification co-developed by Microsoft, Google, GoDaddy, and Hugging Face that defines how agents find tools, skills, and other agents at runtime rather than at install time. It replaces the “hardcode tool locations, ship a static catalog” model with a two-layer approach: a static ai-catalog.json manifest published at a well-known URL by capability providers, plus a dynamic registry API for live semantic search. Hugging Face ships the first reference implementation as a Discover Tool (REST API + MCP endpoint), letting agents query “I need something that does X” in natural language and get back ranked capability matches across federated registries.

Implications

This feeds the native-tooling speed thread and the host-ownership war:

• The install-first bottleneck breaks. Current agent tooling assumes a human (or a DevOps pipeline) pre-configures which tools are available. ARD makes discovery a runtime capability — agents can identify and adopt new tools mid-task. That compresses the latency between “a new capability exists” and “agents are using it” from a deployment cycle to a search query.
• Who owns the registry owns the capability graph. The spec is open, but the reference implementation is Hugging Face’s. Whoever runs the authoritative registry for a domain controls what agents can discover, rank, and route to — a structural position analogous to what app stores hold for mobile software. Microsoft and Google co-authoring the spec while HF runs the implementation is worth watching as adoption grows.
• MCP as capability primitive. ARD’s search results return MCP server endpoints as discoverable artifacts — this is the protocol layer that makes the host-ownership dynamic concrete. An agent finding an MCP server through ARD is an agent whose capability surface just expanded without any human approval step.