On June 17, 2026, the industry shipped a discovery layer for the agent ecosystem. ARD lets AI clients ask what is available for this task and get back verified capabilities — without bloating context windows. Here is the practitioner's breakdown.
The agent stack finally split into two layers everyone can name: execution (MCP, A2A, APIs) and discovery (ARD). Before ARD, agents either hardcoded tool lists — brittle and stale — or pre-loaded every MCP server into context — expensive and noisy. ARD introduces a federated discovery layer where organizations publish ai-catalog.json at their domain and clients search at runtime for matching capabilities.
How ARD works in practice
- Publish ai-catalog.json at /.well-known/ on your domain listing agents, MCP servers, skills, and APIs
- Discovery services index catalogs across federated registries — no single central directory required
- Clients perform intent-based search: given a task, return matching resources with trust metadata
- Selected resources invoke through native protocols — MCP, REST, workflow engines — not through ARD itself
- GitHub agent finder uses ARD so Copilot discovers MCP servers and skills dynamically per prompt
Enterprise implications
Google's Agent Registry in Gemini Enterprise connects internal catalogs to the federated network with governance — trust policies, publisher onboarding, and search ranking under enterprise control. This mirrors how I think about production agent deployments: discovery must be governed, not anarchic. Allowlists, audit trails, and synthetic data disclosure (patterns in my Google ADK Portfolio) become more important when agents can discover new tools at runtime.
Get started
Publish your first catalog following the ARD quickstart at agenticresourcediscovery.org. Read the spec on GitHub (ards-project/ard-spec). For execution-layer reference implementations, see my MCP field guide and Google ADK Portfolio articles on draketalley.ai/blog.
Frequently asked questions
- What is Agentic Resource Discovery (ARD)?
- An open specification (announced June 17, 2026) for publishing, discovering, and verifying AI capabilities — agents, MCP servers, skills, APIs — via standardized ai-catalog.json manifests at /.well-known paths. ARD sits before invocation; resources are called through their native protocol (MCP, REST, etc.).
- How is ARD different from MCP?
- MCP is an execution protocol — how agents invoke tools. ARD is a discovery protocol — how agents find the right tool for a task at runtime. They complement each other: ARD answers what is available; MCP answers how to call it.
- Who supports ARD?
- Working group participants include Google, Microsoft, GitHub, Hugging Face, Nvidia, Salesforce, Snowflake, Databricks, and others. GitHub launched agent finder for Copilot built on ARD — dynamic runtime discovery instead of pre-loading every MCP server into context.