On June 17, 2026, Google published the Agentic Resource Discovery specification alongside ten other major technology companies, solving a problem that has quietly limited every AI agent deployment: agents can only use tools that someone manually wired up for them ahead of time.

The specification, known as ARD, changes that. It gives AI agents a way to search for, find, and verify the capabilities they need at runtime, without a developer having to connect everything by hand.

What ARD Is and Why It Matters

Think of ARD as a discovery layer that sits underneath the protocols businesses are already hearing about. MCP (Model Context Protocol) defines how agents call tools. A2A (Agent-to-Agent) defines how agents talk to each other. ARD solves the problem that comes before both: how does an agent know what tools and other agents exist in the first place?

Before ARD, the answer was simple and limiting. A developer built a list of tools, wired them into the agent, and that was it. If a new tool launched that would have been useful, the agent did not know it existed. Every expansion to an agent’s capabilities required a developer to go in and manually add the connection.

ARD replaces that with a web-scale discovery model. Organizations publish a machine-readable file called ai-catalog.json at a standard location on their domain. This file lists their available AI capabilities: agents, tools, skills, APIs, knowledge bundles, anything they want other agents to be able to find and use. Registries then crawl these catalog files, index them, and respond to natural-language queries from agents looking for specific capabilities.

The result is that an agent can ask “what tool helps me process a PDF invoice?” and get back a ranked list of verified options, then connect to the right one, all at runtime.

Who Is Behind It

ARD was co-authored by Junjie Bu at Google, R.V. Guha at Microsoft, and Shaun Smith at Hugging Face. The specification is licensed under Apache 2.0 and builds on foundational work from the Linux Foundation’s AI Catalog Working Group.

The eleven launch contributors span nearly every major player in enterprise software: Cisco, Databricks, GitHub, GoDaddy, Google, Hugging Face, Microsoft, NVIDIA, Salesforce, ServiceNow, and Snowflake.

The model is federated, meaning no single company controls the catalog. Any organization can publish their own catalog and run their own registry. Registries can cross-reference each other without a central authority.

What Is Already Built on It

GitHub launched Agent Finder on the same day as the ARD specification. Built directly on ARD, it lets GitHub Copilot dynamically find and call MCP servers, skills, tools, and agents for a given task at runtime, without developers needing to preconnect everything. Developers and enterprises can control which AI resources their agents can access.

Hugging Face launched a Discover Tool that indexes its catalog of Spaces, Skills, and MCP servers in the ARD format, making thousands of AI capabilities searchable through the specification from day one.

What This Means for Business

Agent capability gaps get smaller, automatically. Right now, businesses that deploy AI agents often hit a ceiling: the agent can only do what it was initially configured to do. ARD creates a path toward agents that extend their own capabilities by finding what they need when they need it. That reduces ongoing maintenance costs and the lag between new tools becoming available and agents being able to use them.

The open standard prevents lock-in. Having Google, Microsoft, Cisco, Salesforce, and ServiceNow all agree on the same discovery layer is unusual. It means a business can publish its AI capabilities once and have them findable by agents built on any of those platforms. That is a meaningful difference from a world where each vendor builds their own proprietary discovery mechanism.

Governance becomes more important, not less. ARD includes verification mechanisms so agents can confirm a capability is what it claims to be before connecting to it. For businesses concerned about AI agents making unauthorized connections or being misled by malicious tools, that verification layer matters. The ability to control which resources an agent can discover and connect to is a governance primitive that enterprise IT teams will want to understand.

The infrastructure for agentic AI is maturing fast. ARD, alongside MCP and A2A, suggests the industry is converging on a stack: a protocol for discovery, a protocol for tool-calling, and a protocol for agent-to-agent communication. Businesses that are still evaluating whether to invest in AI agents are making that decision in a market where the foundational infrastructure is solidifying around open standards. That reduces long-term technology risk considerably.

For Enterprise DNA clients building out their AI infrastructure, ARD is a signal worth tracking. As the discovery layer becomes standard, the question shifts from “can our agents find the tools they need?” to “which tools do we want our agents to find?”