sierra-research/tau2-bench

τ-Bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains

What it solves

$\tau$-Bench provides a standardized way to evaluate how well AI customer service agents handle real-world interactions. It addresses the difficulty of testing agents that must follow strict policies, use external tools, and interact with users in complex, multi-turn conversations across different modalities (text and voice).

How it works

The framework simulates interactions between an agent and a user simulator. It defines specific "domains" (such as airline, retail, telecom, and banking) where each domain includes a set of operational policies the agent must follow, tools it can call, and a series of tasks to complete.

It supports two primary communication modes:

  • Text (half-duplex): Traditional turn-based chat.
  • Voice (full-duplex): Real-time, simultaneous audio interactions using providers like OpenAI, Gemini, and xAI.

Additionally, it includes a specialized knowledge domain for testing RAG (Retrieval-Augmented Generation) capabilities, allowing agents to search unstructured documents to resolve customer issues.

Who it’s for

  • Agent Developers: Those building conversational AI for customer service who need to verify their agent's adherence to policy and tool-use accuracy.
  • AI Researchers: People studying the reliability of LLM-based agents in tool-heavy, multi-turn environments.
  • RL Practitioners: Developers using the provided Gymnasium-compatible interface to train agents via reinforcement learning.

Highlights

  • Multimodal Evaluation: Supports both text-based and end-to-end audio-native voice interactions.
  • Diverse Real-World Domains: Pre-built scenarios for airline, retail, telecom, and banking industries.
  • Knowledge-Aware Testing: Dedicated RAG pipelines and document search for evaluating knowledge retrieval.
  • Rigorous Grading: Includes a refined set of tasks with over 75 fixes to ensure evaluation accuracy and eliminate ambiguous constraints.