Hugging Face adds search for agents

Hugging Face has introduced Agentic Resource Discovery (ARD), an open draft specification, together with a first Hub implementation called the Discover Tool. The central fact is specific: ARD is meant to give AI agents a search layer for finding tools, skills, applications and other agents at runtime, instead of relying only on connectors that were installed in advance. The post names contributors from Microsoft, Google, GoDaddy, Hugging Face and others, and positions the work next to MCP, Skills and A2A.

The problem is no longer abstract. Agents can call tools through MCP, consume instructions through Skills, or talk to other agents through A2A, but those protocols often assume that a user or developer already knows what should be connected. That model becomes brittle inside a company, where teams have to maintain lists of servers, internal services, models, scripts and access rules. ARD proposes a searchable registry enriched with metadata such as publisher identity, representative queries, compliance attestations and tags.

The useful part of the announcement is the implementation. Hugging Face says its Discover Tool exposes thousands of Skills, machine learning applications and MCP servers as ARD catalog entries. The Hub builds on its existing semantic search across Spaces and agent-compatible applications, then translates those results into the ARD format. In practice, a client can send a natural-language query, receive a ranked list of relevant capabilities, and let the model invoke the selected tool. The choice is not left entirely inside the model: it goes through a more structured discovery step.

The detail also matters for tool publishers. A registry entry can describe not only what a service does, but who publishes it, where it is reliable, and which constraints apply. That is meaningfully different from a general web search, because an agent is choosing an executable capability, not just reading a page. Metadata quality therefore becomes part of both security and product experience.

If the draft matures, the change will be felt mostly in agent plumbing rather than in splashy user-facing claims. An internal assistant could search for “check a supplier invoice” and find an approved tool, instead of asking every team to preinstall the right connector. A developer could discover a useful MCP server without searching a separate directory by hand. Caution is still needed: a dynamic registry raises real questions about security, provenance and governance. But ARD gives a name and a format to a need that is already obvious: moving agents from a fixed catalog of tools toward controlled discovery of the capabilities available around them.