Browser Use adds outcome checks after browser tasks
Browser Use version 1.0.5 sends an additional request to the agent’s own model after each browser task to assess the outcome. The check can record rejected inputs, blocked access and other site problems for subsequent tasks, including successful runs. Users can supply assessment criteria or disable the function; the change is described in the maintainers’ release document.
Artificial Intelligence··Night
The agent’s model checks the result
Browser Use, software for carrying out browser tasks with an AI agent, made version 1.0.5 available on November 14. After a task ends, an additional request asks the same model to assess its outcome. The judgment records success, a failure reason and a description, together with flags for malicious instructions and site access blocks. This account is based on the maintainers’ single release document.[1]
Screenshots and execution history inform the check
The judge receives the execution trace, final page and last screenshots. Disabling vision removes the visual material. Users can supply their own criteria or expected answer, or turn judging off. The release also restores older assessment options. Notes about individual websites can describe rejected form input, errors, blocked access or rate limits, even when the overall task succeeds.[1]
Site problems become notes for later tasks
Maintainers tested 13 real runs, each assessed three times. On runs where a website pushed back, judgments containing lessons increased from six out of 15 to all 15. Success verdicts remained unchanged. The test concerns passing site problems on to subsequent tasks, and the developers conducted these trials themselves. The code is available in the GitHub release; distribution through the Python package index requires separate publishing approval.[1]