Enterprise

Google, Microsoft and nine others launch ARD, an open standard so AI agents can find their own tools

The Agentic Resource Discovery spec, unveiled by an 11-company coalition that conspicuously excludes OpenAI and Anthropic, aims to be a search engine for the agentic enterprise web.

Photo: Unsplash / Conny Schneider — Abstract blue network of glowing nodes and connections representing distributed AI agents discovering capabilities across the web

A coalition of 11 tech companies on June 17 published the Agentic Resource Discovery specification, an Apache 2.0–licensed open standard that lets AI agents locate tools, skills, and other agents across organizational boundaries at runtime. Google and Microsoft are the visible leads. Cisco, Databricks, GitHub, GoDaddy, Hugging Face, Nvidia, Salesforce, ServiceNow, and Snowflake round out the group. OpenAI and Anthropic aren’t on the list.

That absence is the story underneath the story. The Information reports that analysts read the omission as deliberate, with Google and Microsoft building enterprise agentic infrastructure that doesn’t depend on the two frontier-model labs currently setting the consumer narrative. It’s a familiar maneuver: the hyperscalers letting the model labs sprint ahead on capability while quietly pouring the concrete the enterprise will actually run on.

ARD sits one layer above the protocols that already exist. MCP standardized how agents call tools; A2A standardized how agents call other agents. Both assume the agent already knows what it’s looking for. ARD handles the step before invocation, the discovery problem, and according to Google’s announcement it does so through two primitives: catalogs published under an organization’s own domain, with domain ownership serving as the cryptographic root of trust, and registries that act as search engines for the agentic web. The spec is built on the AI Catalog data model maintained under the Linux Foundation’s AI Catalog Working Group.

Two reference implementations shipped alongside it. GitHub’s agent finder gives Copilot the ability to dynamically discover and call MCP servers, skills, tools, and agents at runtime; GitHub says it’s available across all Copilot plans, surfaces resources rather than auto-installing them, and respects managed Copilot settings. Hugging Face’s Discover Tool offers search across thousands of skills, ML applications, and MCP servers on the Hub and other ARD services.

The enterprise pitch is governance. CIO.com notes that ARD embeds SOC2, HIPAA, and GDPR compliance attestations directly into catalog entries, alongside cryptographic identity verification, giving IT a way to govern agent tool access at the infrastructure level rather than inside each application.

Whether OpenAI and Anthropic join, ignore, or counter with their own spec is the variable that decides how open this layer stays.

Sources