EnterpriseClawBench: 852 Real-World Agent Tasks, Best Score Only 0.663

EnterpriseClawBench derives 852 reproducible enterprise agent tasks from real workplace sessions, each with fixtures, role classes, skill subclasses, and semantic rubrics. The best-performing configuration, Codex with GPT-5.5, scores only 0.663, exposing a significant gap in enterprise agent capability. Benchmark data is not released, but the construction and evaluation protocol is the reusable contribution.