Google has published the Agentic Resource Discovery specification, an open standard that tries to solve a plumbing problem holding back AI agents: how an agent finds the right tool or service somewhere on the web, decides which one to use, and verifies it is safe to connect to. Today those capabilities live in scattered, separate registries, and an agent built in one ecosystem has little way to reach a tool hosted in another.
The design is deliberately simple and webby. An organization publishes a static ai-catalog.json file at a well-known path on its domain, listing the AI capabilities it offers, much as websites publish a robots.txt file or a sitemap. A separate registry API then crawls and indexes those published catalogs and answers natural-language queries with ranked matches, so an agent can ask for a capability in plain language and get back candidates it can actually connect to. The specification is licensed under Apache 2.0 and built on a shared data model Google calls the AI Catalog.
It slots in alongside, not against, the protocols already in play. Where something like the Model Context Protocol standardizes how an agent talks to a tool once it has found it, ARD addresses the step before that, discovery and verification: which tools exist, and which can be trusted. The verification angle matters as much as the discovery, because an agent that will autonomously call a service needs some basis for deciding it is the real, safe one and not a lookalike.
The pitch is an open agentic web, a layer where tools and agents from different vendors can find each other without everyone funneling through one company's directory. Whether it takes hold depends on adoption, since a discovery standard is only as useful as the catalogs that publish to it and the agents that read them, and Google is not the only party with a proposal for how the agent ecosystem should interconnect. But the problem it names is real: as agents multiply, the web needs a way for them to find and vet each other, and a robots.txt shaped answer is at least a familiar place to start.