Browser Agent action and safety flowFlowchart showing untrusted page content, model action proposal, content and action safety checks, permission checks, execution or pause, and post-action verification. Browser Agent · Action and safety flow Treat page content as untrusted input; authorize actions separately from interpreting intent Observe pageDOM · AX tree · screenshot Untrusted contentWeb text · email · documents Model reasoningCompare with user goal Action proposalTarget · operation · data Safety screeningContent injection + action risk Authorization policySite permission · user approval · one-way door Pause / ask userBlocked · approval required Execute and verifyRe-read state · record evidence Page content informs reasoning but never becomes an authorization source. A successful tool call is not proof that the business outcome succeeded.

Editorial architecture diagram · Source: public product documentation. Internal vendor implementation details are not implied.