Agentic Resource Discovery is an open specification for publishing, finding, and verifying agents, MCP servers, skills, and APIs across the web.
The web is about to be navigated by far more software agents than people. Yet an agent capable of taking action still faces a basic problem: how does it find the right tool, identify its publisher, and verify that connecting is safe?
On June 17, 2026, Google and industry partners announced Agentic Resource Discovery (ARD), an open specification for publishing, finding, and verifying resources for AI agents. It is less flashy than a new model, but potentially more structural: ARD aims to give the agentic web a shared discovery layer.
Execution exists; discovery is missing
MCP, A2A, OpenAPI, and agent frameworks already support tool invocation and collaboration. What does not scale is assuming that every agent knows every endpoint in advance. An operations agent may need observability, documentation, deployment history, and ticketing capabilities distributed across providers, protocols, and domains.
ARD sits before invocation. It does not execute a tool or replace its protocol. It describes what is available, who publishes it, and how it can be reached. The selected resource is then invoked through its native mechanism.
How ARD works
- Catalogs. An organization publishes an
ai-catalog.jsonfile on its domain, describing agents, MCP servers, skills, APIs, or workflows. - Registries. Federated discovery services index catalogs and answer agent searches using their own trust, ranking, and access policies.
- the provider hosts the catalog at
/.well-known/ai-catalog.json; - a registry indexes it, or an agent reads it directly from a known domain;
- verifiable metadata helps confirm publisher identity and provenance;
- the agent connects directly using MCP, A2A, OpenAPI, or another declared protocol.
ARD is open source under Apache 2.0 and is being developed with participation from organizations including Google, Microsoft, GitHub, Hugging Face, Cisco, NVIDIA, Salesforce, and ServiceNow.
ARD is not another MCP
MCP defines how an AI client communicates with tools and context; A2A supports agent collaboration; ARD helps discover which resource to use. These layers complement one another.
ARD is not a mandatory central catalog either. Public, vertical, and enterprise registries can coexist, each applying different inclusion and trust policies. Identity remains anchored to domains, discovery is federated, and execution stays direct.
From SEO to agent discoverability
Organizations have long optimized pages so search engines can understand their content. ARD introduces a related requirement: make operational capabilities understandable to agents.
A useful catalog needs precise descriptions, representative queries, declared versions, limits, and protocols. Metadata quality becomes part of the product because it affects whether an agent can find a resource and how confidently it can select it.
What enterprises should consider
- The company domain becomes a trust anchor for official capabilities.
- Internal and partner tools become searchable through private registries without being public.
- Governance enters discovery: identity, egress policy, version pinning, and audit should precede execution.
- Integrations become less rigid: agents can discover compatible capabilities at runtime rather than relying only on manually configured connectors.
A cautious starting checklist
- inventory supported agents, APIs, MCP servers, and workflows;
- separate public, partner, and internal resources;
- define ownership, versioning, and lifecycle for every entry;
- publish clear capability and limitation metadata without exposing secrets;
- keep authentication, authorization, logging, and rate limits at the execution layer.
ARD is young and its ecosystem is still forming. It is not a reason to expose infrastructure indiscriminately. It is a signal that open-web discovery for autonomous software is becoming a real engineering concern.
ElaraTeq approaches these standards pragmatically: controlled experimentation, verifiable infrastructure, and security before automation. If you want to assess how your services and data can prepare for the agentic web, let’s talk.
