Sierra, an AI startup, has released a new benchmark called 𝜏-bench (TAU for Tool-Agent-User) to evaluate the performance and reliability of AI agents in real-world settings. The benchmark assesses AI agents' interaction with dynamic users and tools. Initial results indicate that AI agents built with simple LLM constructs, such as function calling or ReAct, perform poorly on complex tasks. The study includes an evaluation of 12 popular LLMs, revealing significant performance gaps in real-world applications and real work.