An AI agent can use almost any capability you hand it — MCP for tools, A2A for other agents, plain HTTP for APIs. What it still can't do is find one it wasn't already given.

Project description

An AI agent can use almost any capability you hand it — MCP for tools, A2A for other agents, plain HTTP for APIs. What it still can't do is find one it wasn't already given. There's no web-scale way for an agent to ask: "What's available for this task, across organizations — and which of these can I actually trust?" Today that gap is filled by walled gardens: one vendor's marketplace, one company's hand-maintained list. An agent on platform A can't discover what's published on platform B without a human wiring them together first. Agentic Resource Discovery (ARD) is a proposed answer, and it borrows the shape of the web itself. It's built from two primitives: Catalogs — a static JSON file you publish under your own domain, at /.well-known/ai-catalog.json. Dumb on purpose, no query logic. Anyone with a domain can publish one, for free. Registries — federated search engines that crawl those catalogs and return ranked results. They refer queries to one another, so discovery isn't trapped in a single index. Catalogs are to websites as registries are to search engines. The part I find sharpest is trust: the domain is the root. If you fetched the catalog over valid TLS, the web's existing PKI already vouches for who served it — no new central registrar required. And ARD is careful to separate two things everyone will be tempted to conflate: a registry's relevance score is not a safety rating. It's complementary to MCP and A2A, not competitive — the front door that gets the address into your hands before they take over. I made a 40-second explainer, plus an opinionated companion repo that distills the v0.9 draft spec and flags where I think it's still soft: signatures are optional, registries are unaudited black boxes, and capability semantics are free text. Repo: github.com/inthepond/ARD #AIAgents #MCP #AgenticAI #OpenStandards