Agentic Resource Discovery: An Open Standard for AI Agent Interoperability
This article introduces the Agentic Resource Discovery (ARD) Specification, an open standard by Google and partners, designed to enable AI agents to publish, discover, and verify external tools and APIs across organizational boundaries. It addresses the critical need for a common mechanism for agents to find and trust resources, complementing existing protocols for tool invocation by focusing on discovery and a governed, secure approach.
Read original on InfoQ ArchitectureThe Challenge of AI Agent Tool Discovery
As AI agents become more sophisticated and prevalent, especially in enterprise environments, their ability to leverage external tools, APIs, and services is paramount. However, a significant architectural challenge arises in how these agents discover and securely integrate with a diverse and distributed landscape of capabilities. Traditional approaches often rely on hardcoded integrations or static lists, which are brittle and don't scale across organizational boundaries or evolving ecosystems.
Agentic Resource Discovery (ARD) Specification
The Agentic Resource Discovery (ARD) Specification proposes an open standard to standardize this discovery process. It acts as a complementary layer to existing execution protocols like the Model Context Protocol (MCP) and OpenAPI, focusing solely on how an agent finds a tool rather than how it invokes it. Key design principles include built-in security, identity, and governance, which are crucial for enterprise adoption where resources are often sensitive and distributed.
Core Constructs: Catalogs and Registries
- Catalogs: Organizations publish a machine-readable `ai-catalog.json` file within their domain. This file describes available capabilities (tools, APIs, skills, agent endpoints) that their agents can offer or use.
- Registries: These aggregate multiple `ai-catalog.json` files. Registries allow AI agents to search for resources based on task intent, abstracting away the need for explicit knowledge of every endpoint. This creates a federated discovery model where agents can locate relevant resources dynamically across different domains.
Trust and Verification
A central tenet of ARD is trust. The specification incorporates domain-based ownership and verification mechanisms. This allows agents to validate the authenticity of discovered resources before establishing connections, mitigating risks associated with autonomous actions across third-party and enterprise systems.
The ARD specification is designed for interoperability and evolution. It includes schemas, trust mechanisms, and guidelines. Early implementations, such as GitHub’s Agent Finder in Copilot and Hugging Face’s Discover Tool, demonstrate its practical applicability in enabling runtime capability discovery for AI agents.