On June 17, 2026, Google announced the Agentic Resource Discovery (ARD) specification. It’s an open standard for publishing, discovering, and verifying AI capabilities across the web. Think of it as DNS for agents.
This isn’t another protocol. It’s the layer that makes all the other protocols work together.
Before yesterday, the agent internet had three major standards, but they didn’t talk to each other:
Each standard solved a piece of the puzzle. But the piece nobody solved: how does an agent find the right tool in the first place?
If an agent needs a weather API, it can’t query every MCP server on the internet. If it needs a payment processor, it can’t scan every llms.txt file. There was no search layer. No discovery layer.
ARD is that layer.
ARD uses two primitives:
ai-catalog.json file you host on your domain. It lists what you offer: MCP servers, A2A agents, OpenAPI tools, or nested catalogs. Because it’s served from your domain, the domain itself becomes the root of trust.{
"name": "Example Corp Agent Hub",
"capabilities": [
{
"type": "mcp",
"name": "Weather MCP Server",
"url": "https://example.com/weather/mcp",
"description": "Real-time weather data for any location"
},
{
"type": "a2a",
"name": "Customer Support Agent",
"url": "https://example.com/support/a2a",
"description": "Handles refunds, returns, and account issues"
},
{
"type": "openapi",
"name": "Inventory REST API",
"url": "https://example.com/inventory/openapi.json",
"description": "Warehouse stock levels and order fulfillment"
}
]
}
The magic: after discovery, ARD steps out of the way. The agent connects directly using whatever native protocol the tool uses — MCP, A2A, OpenAPI, whatever. ARD doesn’t replace MCP. It makes MCP findable.
Google’s demo video shows an agent discovering and executing completely new capabilities at runtime. Four phases:
ai-catalog.json on their domainThis is the moment the agent internet becomes an actual internet. Before ARD, agents could only use tools they were pre-configured with. After ARD, they can discover tools they’ve never heard of.
The complete agent stack now looks like this:
| Layer | Standard | Question It Answers |
|---|---|---|
| Discovery | ARD (Google, June 2026) | Where is the right tool? |
| Site map | llms.txt (Jeremy Howard, Sept 2024) | What’s on this website? |
| Transport | MCP (Anthropic, Nov 2024) | How do I call the tool? |
| Collaboration | A2A (Google, April 2025) | How do agents work together? |
| Specs | .dog/.dag (dotdog, June 2026) | What are the entities and rules? |
Each layer answers a different question. Together, they form a complete stack for agents to navigate the web the way browsers navigate it for humans.
If you build tools agents might use — APIs, MCP servers, A2A agents, anything — you should publish an ai-catalog.json on your domain. It’s a one-time setup. The quickstart guide walks through it.
Once your catalog is up, registries will crawl it. Agents will find your tools. And because ARD includes cryptographic verification, they’ll trust them.
This is the same shift that happened with web search in the 90s. Before Google, you had to know a URL to visit a site. After Google, you could search for what you needed and discover sites you’d never heard of. ARD does this for agents.
ARD solves discovery. But what happens after the agent connects?
It finds a catalog entry: “MCP server at this URL.” It connects. It gets a list of tools. But how does it know what those tools mean? What entities they operate on? What states those entities go through? How they relate to each other?
That’s the next layer. Discovery gets you to the door. Content tells you what’s inside.
Sources: