The rule. Agentic browsers now operate your interface on a person's behalf, so machine legibility (semantic markup, labeled controls, predictable flows) is a UX requirement, not an accessibility bonus.
Why. Agents work an interface in three ways: interpreting screenshots, parsing the browser's accessibility tree, or skipping the UI for an API, and "interfaces that are well-built for accessibility are already more legible to agents: semantic HTML, properly labeled elements, clear roles, and logical page hierarchy also serve agent users" (Gibbons & Moran, 2026). This is shipped software, not a forecast: Claude in Chrome performs actions like "clicking, typing, and filling out forms the way a person would" and is available on all paid plans (Anthropic, 2026), and Agent mode in OpenAI's Atlas browser "lets ChatGPT take actions in the browser for you" (OpenAI, n.d.).
Seen in the wild. Atlas pauses "to ensure you're watching" when its agent acts on sensitive sites such as financial institutions (OpenAI, n.d.): the vendors already assume software will drive your UI with a human supervising.
Verified as of 2026-09-27 against current Anthropic and OpenAI documentation.