The Autonomous Web Agent (AWA) 1.5 system has achieved a score of 57.14% on the WebArena benchmark, surpassing the previous record of 35.8%. This result marks the closest any AI system has come to human-level performance, which currently stands at 78%.

  • AWA 1.5 uses custom HTML-DOM parsing and iterative processing to mimic human-like interaction.
  • WebArena provides a controlled environment with containerized sites and unambiguous outputs for reliable evaluation.
  • The agent faces challenges with drag-and-click actions, stateful tasks, and strict string matching requirements.

The achievement demonstrates significant progress in web autonomy by addressing the inherent noise and variability of evaluating agents on the open web.