What the Agentic Resource Discovery Specification Actually Does
There's always been a messy gap in the AI agent world that nobody really talked about — not because it wasn't a problem, but because it's the kind of foundational infrastructure issue that's easy to ignore until suddenly everything breaks at scale. Here's the gap: AI agents are getting really good at doing things, but they have no reliable way to find the tools they need to do those things. Each connection has to be hardcoded. Every integration is manual. And across organizational boundaries? Forget it.
That's the problem the Agentic Resource Discovery specification — ARD for short — is designed to fix. Published on June 17 by a coalition of eleven major technology companies, ARD is an open protocol that gives AI agents a standardized way to discover, evaluate, and connect to tools and services across the internet. No hardcoded integrations required.
Think of it this way: right now, an AI agent is a little like a contractor who shows up to a job site but has to personally know every supplier in town before they can get materials. ARD is basically a trade directory — a shared, searchable catalog that lets the contractor find what they need, from whoever has it, without needing a personal introduction first.
The Coalition Behind It — and Who's Notably Missing
Google and Microsoft led the initiative, but they didn't go it alone. Cisco, Databricks, GitHub, GoDaddy, Hugging Face, Nvidia, Salesforce, ServiceNow, and Snowflake all signed on as launch partners — eleven organizations in total. The spec is licensed under Apache 2.0, which means it's genuinely open, and it builds on the AI Catalog data model from the Linux Foundation's AI Catalog Working Group.
But honestly, the names that aren't on that list are just as interesting as the ones that are. OpenAI — which built out the Model Context Protocol competitor ecosystem — isn't there. And Anthropic, the company that actually originated MCP, is absent too. That's not a minor footnote. It means the two most prominent AI labs in the world sat out the launch of what's being positioned as a foundational layer for the agentic web.
ARD is careful to frame itself not as a replacement for MCP or Google's Agent-to-Agent protocol, but as a discovery layer that sits above them. The idea is that ARD helps an agent decide which capability to use before it ever invokes anything. Once it finds what it needs, it connects using whatever native protocol that resource supports — MCP, a standard API, or something else entirely.
Catalogs and Registries: The Two Core Primitives
The spec is built around two fundamental building blocks, and once you understand them, the whole architecture clicks into place.
The first is the catalog. Organizations publish an ai-catalog.json manifest file at a predictable, well-known path on their domain. This file describes what capabilities that organization makes available — think of it as a machine-readable menu of what an AI agent can do with your services.
The second is the registry. Registries crawl and index those catalogs across the web, then return ranked matches when an agent sends a natural-language discovery query. So instead of an agent knowing ahead of time that a specific tool exists at a specific endpoint, it can simply ask — in plain language — "what can help me do X?" and get back a ranked list of relevant capabilities.
Google engineers Junjie Bu and Srinivas Krishnan put it plainly in their announcement: the missing piece wasn't the protocols themselves, but a standard way for agents to discover capabilities across organizational boundaries and establish trust in what they find. ARD is that missing piece.
How Google and Microsoft Are Building It Into Their Products
This isn't just a paper spec — both Google and Microsoft are already weaving ARD into their enterprise platforms.
Google is integrating ARD into its Gemini Enterprise Agent Platform through a product called Agent Registry. It goes beyond basic discovery, adding enterprise governance features like globally unique namespaced URNs, egress policies, and cryptographic identity verification. For large organizations managing dozens of internal agents and external integrations, that kind of trust infrastructure isn't optional — it's the whole ballgame.
Microsoft framed ARD as giving AI clients the ability to "tap into new capabilities automatically," which is a fairly understated way to describe what's actually a significant shift in how enterprise AI tools get configured and connected.
GitHub — which Microsoft owns — simultaneously launched Agent Finder for GitHub Copilot. It uses ARD-style discovery indexes to let Copilot locate the right capability for a task without any manual configuration on the user's end. That's the practical, day-to-day version of what ARD enables: your coding assistant just finds what it needs, rather than waiting for someone to wire it up.
Why This Matters for Enterprise AI Deployments
Snowflake probably said it most clearly: ARD addresses "a real and growing issue" — specifically, how to discover all the agents available to an enterprise user through whatever interface they're using. That's a genuinely hard problem at scale.
Right now, most enterprise AI deployments are a patchwork. Different teams have different tools, different agents, different integrations — and getting them to talk to each other requires a lot of custom plumbing. ARD offers a standardized layer that, at least in theory, lets agents navigate that complexity on their own.
The protocol-agnostic design is a smart choice. By letting ARD handle discovery while leaving the actual connection to native protocols, it avoids the trap of trying to replace everything at once. Organizations don't have to rip out existing infrastructure — they just add the catalog manifest and let registries do the indexing.