Agentic Resource Discovery (ARD) is an open specification for publishing, searching, and verifying agent capabilities across organizations and platforms. A “resource” can be an MCP server, an A2A agent card, an Agent Skill, an API, or another callable service.
The One-Sentence Definition
ARD aims to give agents a federated capability search layer: find where a capability lives, understand what it is, verify whether it should be trusted, and learn how to connect.
Many agents today receive a manually configured tool list at deployment time. That works for a handful of resources. Once an organization owns hundreds of agents, Skills, and MCP servers across internal and external catalogs, developers struggle to see what already exists and rebuild the same capability.
Why It Is Entering the Spotlight Now
Participants including Google, Microsoft, and Hugging Face announced ARD in June 2026 as a way to publish and discover agent capabilities across the web. On August 24, AWS announced that its Agent Registry uses ARD for cross-environment discovery and governance, bringing the still-draft specification into a concrete enterprise deployment story.
Its rise resembles Agent Plugins but happens at another stage. Agent Plugins answers “how should a capability be packaged for installation?” ARD answers “which packages, tools, or agents exist, and where can they be found?”
The Core Ideas
First, ARD is federated. It does not require every resource to enter one central marketplace. Organizations can publish resource descriptions under their own domains and catalogs, while discovery services search across them. Domain anchoring connects publisher identity to existing web and organizational boundaries.
Second, it is protocol-neutral. A catalog can describe an MCP tool, A2A agent, Skill, or normal API without converting them into one execution protocol. Third, records are machine-readable: resource type, endpoint, capability, authentication, and metadata need a shared shape before an agent or registry can filter them automatically.
Fourth, discovery is separate from invocation. Finding a resource only establishes that a candidate exists. It does not grant permission to connect or prove safety. Clients still need identity, signature, provenance, authorization, and policy checks.
How It Differs from MCP, A2A, and an Agent Registry
MCP defines how an AI application connects to tools and data. A2A defines communication between independent agents. ARD finds those endpoints before a connection begins. ARD is the directory and search layer; MCP or A2A is the protocol used when the conversation starts.
An Agent Registry is a concrete product or service that stores an organization’s agents and tools. ARD is an open description and discovery contract that can let registries and catalogs understand one another. A registry may implement ARD or remain an internal private catalog.
What It Solves—and What It Does Not
ARD can reduce manual configuration and duplicate development, let agents search for capabilities by task, and help platform teams see what already exists. In an enterprise, discovery records can also link to owners, governance state, and audit information.
It does not automatically determine which tool is best or guarantee that a catalog claim is true. Ranking, reputation, malicious-resource detection, authorization, and runtime isolation remain separate problems. Dynamic discovery also expands the attack surface: if an agent automatically connects to unfamiliar resources, prompt injection and supply-chain risk grow together.
When It Matters
A manual allowlist is simpler for a team with ten fixed tools. ARD becomes useful when resources are spread across teams, clouds, and protocols, and lack of visibility is already causing cost or duplication.
As of August 2026, the specification remains v0.9 Draft. It is suitable for pilots, catalog mapping, and compatibility work, but unstable fields should not be treated as a permanent production contract. Implementations should retain versioning, caching, and fallback paths, and require an explicit trust policy before a newly discovered resource is called for the first time.