Browser Agent Faults Slip Past Outcome Checks

Why this matters

Han's results are a warning shot for anyone shipping browser agents on the assumption that "the agent checked its work" means the work got checked. A coordinate bug that lands clicks three-quarters off target sounds like a rounding error, but it's exactly the kind of failure that slips past generic verification, since the URL changes, the page state shifts, and the agent reports success anyway. Worse, the paper shows that tightening the check isn't free: a verifier strict enough to catch half the bad clicks starts flagging false alarms on legitimate actions inside embedded pages, like iframes, where the simple check doesn't know where to look.

For teams building or buying agent tooling, this says the verification layer needs to be judged by the same rigor as the action layer, not rubber-stamped because it exists. "No detectable change" from a tool is not evidence. The real test is whether the system checks the exact outcome it was asked to produce, field text, button state, not just whether something, anything, moved on screen.