ThinkingBox-Bench: 67% of failed agent runs ended cleanly — no tool error, wrong database state

507 stateful business workflows × 20 runs, graded on backend state. Best pass@1: Claude Opus 5.5 at 67.16%; Kimi-K3 solves 93.9% of tasks at least once but only 13.4% all 20 times. Code MIT, data on Hugging Face.