Announcing the Agentic Resource Discovery specification Google Developers Blog
Justy and Cody chat about Google’s new Agentic Resource Discovery (ARD) spec, dissect its core claim, examine the technical trade‑offs, and wonder who should actually care about a universal catalog for AI tools.
Transcript
Justy So, I just read the new Agentic Resource Discovery spec—Google's attempt to give agents a universal address book for tools, skills, and other agents.
Cody Right.
Justy The core claim is simple: without a shared discovery layer, agents get stuck in siloed toolsets, and the web can't scale as a true agentic ecosystem. They argue that three questions—where does the capability live, which one should I use, and how do I verify it—are currently unanswered across org boundaries.
Cody Yeah.
Justy ARD introduces two primitives, catalogs and registries. A catalog is an ai‑catalog.json file served under your domain, and a registry crawls those catalogs, indexes intents, and returns verified capability metadata.
Justy I swear the office coffee tasted like it was brewed with a spare GPU heat sink this morning, so I grabbed an espresso from the corner shop instead.
Cody Mm-hm.
Cody My DC to‑DC train was delayed again, which gave me extra time to think about how a registry would handle rate‑limiting when thousands of agents query it simultaneously.
Justy The blog shows a terminal demo where an ops agent publishes a catalog with entries for an MCP server, an A2A troubleshooting bot, and an OpenAPI wrapper. The agent then queries a registry with a plain‑language intent like ‘investigate latency spike’, and the registry returns the matching MCP endpoint plus a signed JWT for trust. The demo also highlights that the catalog lives at a well‑known path, making discovery as simple as an HTTP GET.
Cody Exactly.
Cody From a systems view, tying identity to domain ownership is clever, but it assumes DNS can't be hijacked and that every provider can manage private keys for signatures. If a malicious actor gains control of the domain, the whole TRUST chain collapses. Also, the spec leaves the capability schema loosely defined, which could force agents to write custom adapters anyway.
Justy Right, right.
Justy For product teams, that means you could expose your internal observability API as an ARD catalog and let any approved agent call it without writing a bespoke connector. It turns a siloed microservice into a first‑class citizen of the agentic web, potentially cutting integration weeks down to a day.
Cody Sure.
Cody But the verification metadata adds network round‑trips and signature checks, which may be non‑trivial in low‑latency incident response. Plus, a global registry indexing every catalog raises privacy concerns—who gets to see which internal tools exist?
Justy Right.
Justy Honestly, the name 'Agentic Resource Discovery' sounds like something straight out of a 90s sci‑fi manual—'Chapter 7: Discovering the Galactic Toolkit.' If only my coffee machine could answer that kind of query!
Cody Okay, okay, but imagine a future where your espresso maker publishes a catalog and your kitchen robot asks it for a double shot. That’d be… oddly efficient.
Justy Anyway, that’s the gist of the spec—something to keep on your radar if you’re building agent‑centric products. Catch you later, Cody.