Google published the Agentic Resource Discovery (ARD) specification on June 17, alongside launch partners including Cisco, Databricks, GitHub, GoDaddy, Hugging Face, Microsoft, Nvidia, Salesforce, ServiceNow, and Snowflake. The spec targets a gap that sits before any tool call happens: an agent has to find the right capability, decide which instance to trust, and confirm it's safe to connect to, all before invoking MCP, A2A, or whatever native protocol actually does the work. The mechanism is deliberately simple. An organization publishes a static ai-catalog.json manifest at a well-known path on its own domain, describing the tools, APIs, skills, and agent endpoints it offers. Separate registry services crawl and index these catalogs, then answer natural-language discovery queries with ranked matches. Because catalogs live under an organization's own domain, domain ownership itself becomes the root of trust — the same logic DNS-based verification has used for decades, just applied to agent capabilities instead of mail servers or TLS certificates. What makes this notable isn't the manifest format — it's the explicit refusal to be anything more than a discovery layer. Google's own framing draws a hard line: ARD finds the resource, then steps out of the way, handing off verifiable trust metadata so the agent can connect using the tool's native protocol. Hugging Face's post on the spec makes the same point — ARD sits entirely before invocation and isn't a replacement for MCP, A2A, or Skills. Two early implementations show what this looks like in practice. Hugging Face launched a Discover Tool wrapping the Hub's semantic search over Spaces, Skills, and MCP servers inside the ARD envelope, making thousands of existing capabilities searchable through the new spec without re-registering anything. GitHub, meanwhile, is shipping an agent finder built on ARD that lets Copilot dynamically discover and call the right MCP servers, skills, and agents at runtime — with the added benefit of keeping unused capabilities out of the context window. Google Cloud's own Agent Registry adds enterprise governance on top: namespaced URNs, egress policies, and cryptographic trust manifests. The spec is still young — a working group draft, not a ratified standard, and the public site is explicit that many discovery services will coexist rather than one central registry winning out, similar to how every enterprise runs its own intranet search rather than depending on a single index. But the domain-as-trust-root design choice is the part worth watching: it means agent discovery inherits the existing security assumptions of the web's DNS and TLS infrastructure rather than inventing a parallel trust system from scratch.