Microsoft and Hugging Face newly spotlighted ThinkingBox as an executable, open benchmark for stateful enterprise-agent outcomes: 507 workflows across five domains, 12 models, and 20 repeated trials per task, scoring terminal backend state and side effects rather than tool-call traces or self-reports. In the vendor team’s reported 121,680 valid trials, 79,853 failed checks; 67.24% of failures nevertheless ended cleanly after a state-changing call with no final tool error. The underlying paper dates to August, so this is a newly surfaced October usability/reproducibility push with Hugging Face/OpenEnv instructions and public repos, not a new benchmark result; figures are author-reported. Sources: https://huggingface.co/blog/microsoft/thinkingbox ; https://arxiv.org/abs/2608.19741 ; https://github.com/microsoft/thinkingbox ; https://github.com/microsoft/thinkingbox-data