Google invented the sitemap. In 2005 it published its own protocol for listing a site's pages; a year later it brought in Yahoo and Microsoft to co-sign it as sitemaps.org, an open standard everyone still uses. In 2011, same play with schema.org: teach the web to structure itself so search can parse it better. Also still the standard. In 2015, AMP tried the same move for mobile pages—except this one routed traffic through Google's own carousel, and was eventually walked back. Same instinct, three tries, two different kinds of endings.
Watching the recently shipped Agentic Resource Discovery spec (ARD) that Google co-authored with Microsoft, GoDaddy and Hugging Face among the contributors, I can't but see it as the latest chapter of that same mission—organize the web for whoever's visiting it, whether that's a crawler, a search engine, or now, an agent.
ARD looks like an open spec for how AI agents find tools and other agents across the web, not just call the ones they already know about. If MCP solved "how do I use this," then ARD is going after "how do I find this," at a scale no one can hardcode by hand anymore.
The mechanics follows the sitemap idea, rebuilt for a web where the visitors are agents: publish a file on your own domain listing what you offer, a registry crawls and indexes it, agents query the registry instead of needing a hardcoded URL. The base layer of trust is domain ownership—the file lives on your domain, so it's presumed to be you—with an optional cryptographic signature on top.
Interestingly, Google didn't build the first implementation this time. GitHub shipped Agent Finder in Copilot the same day as the spec while Hugging Face shipped its own Discover Tool on the Hub. Which is really the same playbook Google itself has been running since 2005: ship early, and let existing distribution quietly turn it into the default.
Whether ARD stays federated in practice is the thing to watch.
More broadly, here is the question I keep turning over with my teams on most of these emerging specs: is this like the early structural shift schema.org turned out to be, or like the version of AMP that promised so much but eventually got walked back? Move too early and you're implementing something that doesn't last. Move too late and you're the last one still hardcoding URLs.
