Agentic Resource Discovery: Let agents search

Contributors across the technology industry, including Microsoft, Google, GoDaddy, and Hugging Face, have developed Agentic Resource Discovery, a draft open specification designed to let AI agents search for capabilities at runtime rather than relying on static, pre-configured installations. Current agent protocols like MCP, Skills, and A2A assume users manually discover and integrate necessary tools. ARD introduces a discovery layer that sits in front of these protocols, allowing clients to search federated registries using natural language and enabling models to invoke returned capabilities dynamically. Hugging Face has released its Discover Tool as a reference implementation of the ARD specification, providing search access to thousands of skills, ML applications, and Model Context Protocol servers across the Hugging Face Hub and other ARD services. The implementation combines the Hub's existing semantic search over Spaces with agent-oriented metadata, translating searches into ARD catalog entries. Filters ensure responses include only running spaces and match requested media types, such as transforming spaces with agents.md files into loadable skills or generating catalog entries for Gradio MCP endpoints over HTTP transport. ARD separates discovery from execution using a REST API and static manifest formats driven by media type, allowing multiple artifact protocols to use the same envelope. Future development plans for Hugging Face include tighter integration with the specification's federation modes and native support for static catalog manifests on user and organization profiles, enabling publishers to advertise capabilities through standard well-known URIs.