代理標準與開發平台
AWS Introduces the ARD Federated Discovery Specification: Let Agents Search for Tools Instead of Stuffing Every Schema into Context
ARD uses a unified REST search interface and `ai-catalog.json` to describe agents, MCP servers, Skills, APIs, and workflows. AWS positions it as a cross-cloud interoperability layer for Agent Registry. The specification remains a v0.9 Proposal, and identity, ranking, and actual authorization are not standardized by the discovery layer.
On August 24, AWS publicly introduced Agentic Resource Discovery (ARD), an effort to address how capabilities can be found as agent systems scale. Current tool-selection approaches typically place a collection of tool descriptions and schemas into the model context. Once the number grows into the thousands, this not only consumes tokens but also makes selection quality sensitive to prompt length. ARD moves candidate filtering into a separate search service, so the model receives only a small set of results relevant to the current task.
The specification publishes a resource catalog at `/.well-known/ai-catalog.json`, which can describe MCP servers, A2A agents, Skills, APIs, datasets, or workflows. Each record uses a domain-anchored `urn:air:` identifier, distinguishes resource formats using IANA media types, and provides the full description through either `url` or inline `data`. The required `POST /search` endpoint accepts task queries, while `POST /explore` and `GET /agents` are optional. `representativeQueries` can help a registry build a semantic index, and `trustManifest` can reference SPIFFE identities, DIDs, attestations, provenance chains, and JWS signatures.
Technically, ARD handles discovery only. It does not replace MCP or A2A invocation protocols, nor does it perform resource authentication, authorization, or execution. AWS says its Agent Registry can use IAM or enterprise JWTs to manage records, approvals, and search. ARD could allow registries across different clouds and internal environments to index one another through a common format, without requiring a separate bilateral connector for every pairing.
Engineering teams should not yet treat it as a finalized standard: the public documentation still labels it a v0.9 Proposal, some media types have yet to be formally registered, and key trust fields remain optional. The next things to watch are whether AWS provides verifiable native ARD import/export support, cross-registry pagination and revocation semantics, and interoperable search ranking and permission filtering across vendors.