The framework is even more “try-hard” than the model—Qwenpaw’s 76.4 score really shows that engineering governance is the true gate for an Agent to be successfully deployed.

Alibaba Tongyi Laboratory has released PawBench v1.0, integrating the base model and execution framework into a single benchmark. Covering 9 major models, it conducts cross-tests on the Hermes, Openclaw, and Qwenpaw frameworks, including 150 tasks and 4,050 test cases. The results show that framework design directly affects an agent’s ability to be deployed in practice: Qwenpaw 76.4, Openclaw 75.4, and Hermes 70.4. Even small models can “beat the bigger players” under the best frameworks. The benchmark proposes four principles: full disclosure, equipping as needed, proactive monitoring, and elastic recovery, recommending that engineering governance be used to unlock the base model’s capabilities.