The Agentic Yellow Pages

A new open specification wants to do for agent tools what search engines did for web pages.

Eleven companies published the Agentic Resource Discovery specification on June 17, proposing a standard way for AI agents to locate tools, skills, and other agents at runtime. The v0.9 draft, released under an Apache 2.0 license by a coalition including Microsoft, Google, GitHub, Hugging Face, Cisco, NVIDIA, and Salesforce, shipped alongside working implementations from GitHub and Hugging Face. The spec builds on the AI Catalog data model under the Linux Foundation, the same organization that took stewardship of MCP earlier this year when Anthropic donated the protocol to the newly formed Agentic AI Foundation. ARD targets the gap between the protocols agents already have for invoking capabilities and the absent standard for finding them.

The missing catalog

MCP standardizes how an agent calls a tool, A2A standardizes delegation between agents, and Skills standardize how an agent consumes instructions. The three protocols collectively handle invocation, the mechanics of executing a capability once an agent has decided to use it. All three assume that the agent already knows which capability it needs and where to find it. Discovery, the step of locating a relevant capability before invoking it, has fallen outside every existing standard.

Manual wiring worked when agents connected to a handful of well-known tools hardcoded into configuration files or plugin stores. The ecosystem now numbers in the hundreds of thousands of available MCP servers, agents, and APIs, with every company and developer able to publish capabilities of their own. The common workaround, loading every available tool description into the LLM’s context window and letting the model pick, hits a hard ceiling at a few dozen options and degrades selection quality well before that limit. Orders of magnitude separate what any single agent can see from what actually exists.

Roadmap to discovery

ARD introduces a static manifest called ai-catalog.json, hosted at a well-known path on the publisher’s domain, that lists the tools, agents, MCP servers, or APIs the organization makes available. The format accommodates any invocation protocol, and an entry can describe an MCP server, an A2A agent, an OpenAPI endpoint, or a nested catalog linking to a department’s own feed. Each entry carries structured metadata describing what the resource does, what inputs it accepts, and what protocol it speaks, alongside a trust manifest that lets the consuming agent verify the publisher’s identity before connecting. Domain ownership serves as the trust anchor because the entity that controls the catalog URL is the entity asserting responsibility for the resources it advertises.

Registries crawl and index these published catalogs, functioning as search engines for agent capabilities. Microsoft’s framing draws an explicit comparison to the early web, when millions of pages existed but most users visited only the sites pre-installed in their browser’s bookmarks. Search engines solved that problem by building a discovery layer that reached everything automatically. ARD proposes to do the same for tools and agents. When an agent needs a capability it lacks, it submits a natural-language description of the task to a registry, which returns ranked matches with the metadata required to verify the publisher and establish a connection. The agent then connects through the selected resource’s own protocol, and ARD plays no further role.

Who will be the early birds?

The spec launched on June 17 with eleven contributors, including Microsoft, Google, Cisco, Databricks, NVIDIA, Salesforce, ServiceNow, and Snowflake, and with reference implementations from two of them. GitHub’s Agent Finder, available on all Copilot plans, searches a curated public catalog or an organization’s private registry and returns ranked MCP servers, skills, and agents that Copilot can load on demand, with enterprise administrators controlling which resources agents may discover. Hugging Face’s Discover Tool wraps the Hub’s existing index of Spaces, Skills, and MCP servers in the ARD envelope, exposing them through both a REST API and a command-line interface. Cisco tied the spec to its AGNTCY Agent Directory, an open-source project under the Linux Foundation, while Google announced that its Gemini Enterprise Agent Platform will support ARD natively in the coming months.

The contributor list spans most of the companies building agent infrastructure. An independent census conducted the day after launch, however, probed 39 domains, including all eleven contributors, and found none serving a discoverable ai-catalog.json at the well-known path the specification defines. ARD at v0.9 has strong institutional backing and no measured catalog adoption outside the two reference implementations. The spec’s value depends on whether the publishers of tools, agents, and APIs adopt the catalog format, because without catalogs to crawl, registries have nothing to index and agents have nothing to discover.

Time will tell

The breadth of the coalition, competitors like Microsoft and Google co-authoring a shared standard, reduces the risk of competing proprietary discovery layers fragmenting the agent ecosystem before it consolidates.

ARD’s authors have produced a spec and two working implementations. Converting a specification into infrastructure requires the unglamorous work of persuading thousands of tool providers to host a JSON file, and that effort remains ahead of the coalition.