Tau-Tau-Bench leaves agents at 23.9%

Tau-Tau-Bench gives coding agents 53 end-to-end customer-service construction tasks, and its strongest tested setup—Claude Opus 5 under Claude Code—passed only 23.9% of held-out simulations against an 82.2% expert-authored reference ceiling, the new paper reports. That gap says agent builders still need human ownership of records, requirements, architecture, and serving budgets; the case for hard denies in Claude Code is a useful companion when a generated agent moves from demo to production.