Open ARD Specification Standardizes AI Agent Tool Discovery

On June 17, 2026, a coalition including Google, Microsoft, and ten other industry partners published the Agentic Resource Discovery (ARD) specification. The standard provides a decentralized discovery layer for the agentic web. Instead of hard-coding integrations, developers can now build AI agents that search, verify, and connect to remote tools and services dynamically at runtime.

Discovery and Trust Primitives

ARD standardizes how capabilities are published across federated registries. Organizations group their tools, skills, and agents into catalogs published under their domain name. Registries then index these catalogs, providing intent-based search endpoints for agents to query based on required capabilities.

The specification relies on the Agent Name Service (ANS) to handle verifiable identity. ARD embeds trust manifests and cryptographic metadata into the discovery process. Before an agent connects to a discovered tool via an execution standard like the Model Context Protocol, it can verify the safety and domain identity of the resource. The entire specification operates under an Apache 2.0 license and uses the Linux Foundation’s AI Catalog data model.

Initial Vendor Implementations

Major platform providers shipped ARD integrations alongside the specification release. Google embedded ARD into its Gemini Enterprise Agent Platform through a new Agent Registry. This implementation provides hosted discovery along with governance controls like egress policies and namespaced URNs. If you implement multi-agent coordination patterns, this centralizes capability routing for enterprise deployments.

Microsoft introduced agent finder for GitHub Copilot. The tool removes the need for manual wire-ups by allowing Copilot to search ARD-compliant registries for relevant MCP servers dynamically. The feature is available across all GitHub Copilot plans.

Hugging Face published an implementation that moves tool selection out of the language model’s context window. Models can invoke a REST endpoint on a registry to retrieve relevant resources. Snowflake added specification support to Snowflake Intelligence for governed discovery within data environments, while Cisco launched the AGNTCY Agent Directory as an open-source demonstration project under the Linux Foundation.

The Agent Protocol Stack

ARD fills a specific routing gap in the current landscape of AI standards. While existing protocols define how agents communicate and execute tasks, they lack a unified way to locate those capabilities.

• MCP: Defines the standard for calling tools and connecting data sources.
• Skills: Standardizes how agent skills and specialized instructions are consumed.
• A2A (Agent-to-Agent): Governs how agents invoke other agents.

By abstracting discovery into federated registries, ARD enables an interoperable network of agents. Execution protocols can scale beyond siloed vendor ecosystems, allowing systems to locate and utilize capabilities across different organizations.

If you build programmatic agent workflows, review your tool integration strategy. Hard-coded tool lists limit system scalability and require constant manual updates. Begin mapping your static tool lists to ARD-compliant catalogs to enable dynamic capability discovery in your architecture.