Google tells developers to build for agents

Web development is getting a new compatibility target. Not a browser, and not exactly a human either. Google’s web.dev team has published fresh guidance telling (web.dev)or and to build sites those systems can reliably understand and operate. The point is simple — if software is going to browse, compare, fill forms, and com(web.dev) a screen. It has to make its structure and actions legible to machines too. (web.dev) ##(web.dev)e is called “Build agent-friendly websites,” and it frames agents as a new kind of web user. Google says some users are shifting from manual browsing to delegating “goal-oriented journeys” to AI systems that can interpret input, plan, and execute actions on a user’s behalf. That is a bigger statement than it sounds. It means Google is not talking about scraping in the old sense. It is talking about the web being used by software that tries to complete tasks. (web.dev)nswer is blunt. Many sites are built to be visually polished for humans, with hover states, shifting layouts, and motion-heavy interfaces, and Google says that setup is “functionally broken for agents.” That is the core idea here. A design can feel modern to a person and still be hard for an agent to parse, because the useful meaning lives in styling tricks and JavaScript behavior instead of in clear structure. (web.dev) ### How do agents “see” a website? Google breaks i(web.dev)cessibility tree. Screenshots help a vision model guess what is important on the page, but Google says that route is slower and more expensive in tokens. HTML gives the agent the DOM structure and relationships between elements. The accessibility tree is the really important part — Google describes it as a high-fidelity map of roles, names, and states that strips away visual noise and exposes functional intent. (web.dev)use the same practices help both groups. Google’s guidance basically says agent-readiness is downstream of good semantic HTML and accessible design. If a control is really a button, mark it up as a button. If a label belongs to a field, wire it correctly. If state changes matter, expose them clearly. This is less like inventing a weird new AI layer and more like finally paying off old web debts. The accessibility tree already exists for assistive technology — agents just benefit from the same semantic clarity. (web([web.dev) What changes for developers? The shift is from page rendering to action design. Developers now have to think about whether an agent can identify the right control, understand what it will do, and confirm what changed after it acts. A human can recover from ambiguity by eyeballing the screen. An agent needs clearer boundaries. That pushes teams toward stable layouts, explicit semantics, predictable flows, and interfaces whose state can be interpreted without guessing. (web.dev) ### Where is Google trying to take t(web.dev)eb standard and invites developers to sign up for an early preview program. The idea is to move beyond agents poking at pages through screenshots and DOM inference and toward sites exposing structured capabilities directly. Basically, instead of an agent hunting for the “book now” button, the site could declare a callable booking tool. That is a much cleaner contract if it sticks. (web.dev) ### Is this just a Google theory piece(web.dev)round agents across web.dev, Google AI developer docs, and Google Cloud. The company is building tools for developers to create agents, and now it is also telling site owners to prepare for agents as visitors. Put those together and the direction is pretty obvious — Google wants an agent-operable web on both sides of the transaction. (web.dev) ### Bottom line The news is not that AI agents exist. The news is that Google is starting t(web.dev)t happens, semantics stop being nice-to-have hygiene and start looking like infrastructure. (web.dev)