Skip to main content
GUI environments fit AWP naturally: the screen is an observation channel, input devices are action types, and the GUI profile pins the schemas so GUI agents become portable across browser adapters.
  • Channels: screenshot (image/jpeg, latest-wins) plus dom or a11y_tree (text/event+json, keyframe/delta). The accessibility tree is usually the best LLM-facing channel.
  • Actions: pointer_move, pointer_click, key_input, scroll, navigate — viewport-relative coordinates per the profile schemas.
  • Safety: navigate outside an allowlist and any credential-entry class are requires_approval — this is where the approval flow earns its keep.
  • Time model: streaming by default; lockstep against a deterministic page harness is excellent for regression-testing agents.
Reference implementation: browser-adapter.