τ^τ-bench makes agent construction the task, and the best configuration passes 23.9% against an 82.2% expert ceiling.

The benchmark hands a developer agent a real client engagement setup, business records, a requirements-holding client, a production API, an inherited codebase, cost and model limits, then scores it by deploying the customer-service agent it built against held-out simulated users (arXiv 2609.04611). Across 53 tasks in four domains, Claude Opus 5 under Claude Code passes 23.9%. The failure modes are recognizably human: shallow queries instead of deep record comprehension, almost no communication back to the client, too little experimentation with architecture and serving setup.