Google published an open standard on June 17, 2026 called Agentic Resource Discovery, or ARD, that lets AI agents automatically find and verify the tools and other agents they need -- across company boundaries -- rather than being hand-wired to each one. Backed at launch by Microsoft, Nvidia, Salesforce, GitHub, Hugging Face and others, it is an attempt to build something the agent ecosystem has been missing: a directory service for the machine web.
Key facts
- What: an open specification for publishing, discovering, and verifying AI capabilities across the web, published June 17, 2026.
- How: organizations publish an ai-catalog.json manifest; the system is federated, so each org runs its own catalog and no single company owns the registry.
- Who: launch contributors include Google, Microsoft, Nvidia, Salesforce, GitHub, Hugging Face, Cisco, Databricks, GoDaddy, ServiceNow, and Snowflake.
- Source: the Google Developers Blog announcement.
The background: over the last year, AI agents -- programs that use tools to get work done -- exploded, and a standard called the Model Context Protocol emerged for connecting an agent to a given tool. But MCP tells an agent how to talk to a tool once it knows where the tool is. It does not solve the earlier problem: how does an agent discover that a useful tool or another agent even exists, especially one owned by a different company? Today that discovery is mostly manual -- a developer wires each connection by hand. As agent workflows sprawl across organizations, that hand-wiring does not scale.
ARD is Google's proposed fix. Any organization publishes a manifest file -- an ai-catalog.json -- that lists the tools and agents it offers and how to verify them. Agents can then query these catalogs to find what is available for a given task and connect on the fly. Crucially, the design is federated: there is no central registry that one company controls. Each enterprise maintains its own catalog while participating in a broader discovery network, the way the web itself is decentralized.
The cleanest analogy, and the one the standard invites, is DNS -- the internet's phone book. Before DNS, reaching a computer meant knowing its numeric address; DNS let you look up "where is this name" automatically, and the web became navigable. ARD aims to be that lookup layer for agents and tools: instead of being told exactly where every capability lives, an agent can just ask the directory. If MCP is the protocol for making a call, ARD is the phone book for finding the number.
Why it matters: interoperability standards are unglamorous and enormously consequential -- they decide whether a technology becomes a connected ecosystem or a pile of walled gardens. A discovery layer that actually gets adopted would unlock multi-agent workflows that span vendors, letting an agent at one company safely find and use a service at another. The unusually broad backer list is the real signal: getting Google, Microsoft, Nvidia, Salesforce, GitHub, and Hugging Face to co-sign the same spec is rare, and coverage from outlets like Help Net Security framed it as a serious cross-industry bid to define how the agent web is organized.
The honest caveat: a published specification with prominent logos is a starting gun, not a finish line. Standards succeed only when enough parties actually implement them, and the graveyard of well-intentioned interoperability specs is large. Some observers also note the security stakes -- a discovery layer that lets agents auto-connect to tools across organizations is exactly the kind of surface where authentication and verification have to be airtight, or it becomes a new avenue for prompt-injection and abuse. ARD builds verification into the spec, but whether it holds up under adversarial pressure and real-world adoption is the question the next year will answer. For now, it is the most credible attempt yet to give the agent ecosystem a shared address book.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)