Claw-Eval: Towards Trustworthy Evaluation of Autonomous Agents