LLM agents are moving from research prototypes to production, handling tasks such as customer service, dispute adjudication, and internal system operation. Building these agents is increasingly delegated to coding agents, yet existing benchmarks rarely test whether an AI system can deliver a complete agent under real client constraints.
We introduce τ^τ‑bench (pronounced hyper‑tau‑bench), which makes the construction of an agent the task itself. A developer agent receives the business records actually kept, a client with requirements, a production API that must be used, an existing codebase to inherit, and limits on serving cost and model usage – exactly the starting point of a real engagement. From this it must produce a full‑featured customer‑service agent, evaluated by deploying it against held‑out simulated users.
Across 53 tasks in four domains, the strongest setup – Claude Opus 5 under Claude Code – passes only 23.9% of the evaluation simulations, while an expert‑authored reference ceiling reaches 82.2%. The failures mirror those seen by human developers: models issue shallow queries instead of deep record comprehension, communicate almost nothing to the client, and experiment too little with agent architecture and serving spend, often shipping the first design that runs.
τ^τ‑bench aims to turn cooperative agent building into a measurable target, giving coding agents a clear benchmark for improvement.
Review