Accessibility is an agent interface.

Chrome's agentic browsing guidance says agents review the accessibility tree to identify interactive elements. A visible button without a programmatic name can therefore block a screen-reader user and leave an agent guessing about the action.

The tree is not a second website. It is a derived representation of the rendered page. Native elements, correct labels and valid relationships usually produce a stronger result than adding broad ARIA after the interface has already become ambiguous.

Lighthouse highlights names and labels, tree integrity and visibility as agent-centric accessibility signals. These are useful deterministic checks, but they do not prove that a complete customer journey works.

A control can pass a label check and still fail the journey. The name may be technically present but ambiguous among repeated actions. A calendar may expose buttons yet lose the selected date after navigation. A modal may trap focus or hide the final state.

Pair static checks with an agent scenario that has a defined goal, allowed actions and deterministic final-state assertion. Record the accessibility snapshot and the visible state at the point of failure so an engineer can compare both representations.

A named control is not usable when focus cannot reach it, activation behaves differently from the native pattern or a modal returns focus to an unknown location. Test the same journey with keyboard input because it exposes ordering, focus management and composite-widget defects that a static tree snapshot cannot show.

Prefer native HTML behavior. When a custom widget is necessary, follow an established keyboard pattern and verify focus order, visible focus, escape behavior and the selected or expanded state after every action.

Loading, validation, empty, error and success states must remain available to people and agents. A visual spinner without an accessible busy state, an error detached from its field or a toast that disappears before inspection can turn a recoverable issue into a blocked task.

Capture the visible UI and accessibility tree before and after the transition. Assert that the intended object, form value or server state changed exactly once, then verify that a retry cannot create a duplicate action.

Chrome also identifies Cumulative Layout Shift as an agentic browsing signal. An agent may resolve a target from semantics and still click the wrong location if late content moves the interface during the action window.

Track CLS for the page and target movement around important actions. Reserve media dimensions, avoid injecting banners above active controls and make loading states match the final layout.