ClawBench V2: reproducible real-world web-agent evaluation
ClawBench is an open benchmark for evaluating AI agents on real-world online workflows. The current release combines V1 (153 tasks) and V2 (130 tasks): 283 tasks across 163 live websites and 15 life categories. Runs use isolated containers, safe submission-request interception, session replay, action screenshots, HTTP traffic, browser actions, agent messages, and a judge-assisted outcome rubric.
Resources:
• Paper: https://arxiv.org/abs/2604.08523
• Code and task definitions: https://github.com/TIGER-AI-Lab/ClawBench
• Project page and leaderboard: https://claw-bench.com/
• Dataset card: https://huggingface.co/datasets/TIGER-Lab/ClawBench
This post is a maintainer-disclosed release note for discoverability and reproducibility. Results should be interpreted with the task version and harness configuration recorded.