ARD fills the gap in agent tooling that MCP left behind

Integrating an AI agent with a single tool is simple enough, but organizations with thousands of resources scattered across multiple cloud providers and internal systems face a thornier challenge: how do agents find what they need? That's where Agentic Resource Discovery (ARD) comes in.

The discovery gap#

The Model Context Protocol (MCP) has become the standard way for AI applications to interface with external tools and data sources. But MCP assumes the client already knows which server it wants to connect to—a limitation that breaks down at scale. As enterprises split infrastructure across cloud platforms, SaaS vendors, and on-premises systems, the overhead of manually configuring every agent-to-tool connection becomes unmanageable.

ARD addresses this missing piece. The specification, which AWS compared to "DNS, but for agents," enables AI clients to search across multiple registries to locate the resources they need, rather than requiring developers to hardcode every connection beforehand.

A collaborative foundation#

The specification was authored by engineers from Google, Microsoft, and Hugging Face—Junjie Bu, R.V. Guha, and Shaun Smith, respectively—and released under Apache 2.0. Contributors from companies including Cisco, Databricks, GitHub, Nvidia, Salesforce, ServiceNow, and Snowflake have shaped its development. AWS has provided feedback and is exploring integration with its own Agent Registry, though the technology itself is not AWS-owned.

How it works#

ARD uses JSON-LD and REST endpoints to enable federated discovery. The specification's core feature is a POST /search endpoint that queries by task, while optional endpoints let clients browse available resources. Crucially, each discovery service can establish its own governance rules—deciding what it returns and which sources it trusts.

This architecture allows organizations to maintain control over their own catalogs and policies while participating in a broader ecosystem. An enterprise might keep internal resources private, for instance, while gaining access to vetted external catalogs when needed—a model AWS terms "describe once, discover everywhere."

The timing imperative#

Agent Registry already provides AWS customers a centralized view of their infrastructure. By connecting with ARD-compatible systems, it could bring external resources into that view without forcing companies to register everything with AWS. The governance model is still taking shape—board structure, membership terms, and potential migration to a neutral body like the W3C or an AI foundation remain to be determined.