Agent-building benchmark scores Claude Opus 5 at 23.9%

τ^τ-bench scores AI coding agents by deploying what they build against simulated users. Per the τ^τ-bench paper, the setup gives each agent business records, a client who holds the requirements, a production API, and limits on serving cost.

The gap matters for anyone building agents to replace a consultant's client work. The top setup still fails most of the simulated interactions it's tested against.

Claude Opus 5 running under Claude Code, the strongest configuration tested, passes 23.9% of the evaluation simulations across 53 tasks in four domains. An expert-authored reference solution passes 82.2% on the same tasks. The researchers trace most of the shortfall to habits, not code quality. Agents ran shallow queries against the client's records instead of reading them deeply. They sent almost no communication back to the client, and experimented too little with their own agent architecture or serving setup.

These agents already write code that runs. What decides pass or fail is whether they ask the client a question before guessing.