Monitoring Web Agents Through Observable Trajectories

Summary

The paper presents a method for predicting web-agent risk from observable execution trajectories rather than internal model signals. Macro and micro features identify whether an agent remains on track, while key-step supervision marks the first uncorrected error associated with eventual failure. Tests on WebArena-Lite and Online Mind2Web across five model backbones show competitive performance, early-intervention capability, and transfer across website categories.