τ-Bench: Tool-Agent Evaluation Framework

Understand the project

Practical overview

Use τ-Bench when you want to compare or inspect agents in simulated customer-service tasks across its listed domains, in text or voice mode.

Before you choose

Core installation covers text-mode mock, airline, retail and telecom domains; banking-knowledge and voice features are optional extras.
The documented setup requires Python 3.12 or later and below 3.14, uv, and configured model API keys.
For voice features, the README identifies PortAudio and FFmpeg as system dependencies on macOS.
The repository warns that banking-knowledge scores from versions before 1.0.1 are not comparable with scores from 1.0.1 or later.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Set up the core text-mode project and launch the documented sample evaluation.

Clone the tau2-bench repository and change into its directory.
Run the command below to install the core text-mode dependencies.
Documented command
```
uv sync
```
Copy .env.example to .env and add the required API keys.
Run tau2 run --domain airline --agent-llm gpt-4.1 --user-llm gpt-4.1 --num-trials 1 --num-tasks 5.

The evaluation results are saved in `data/simulations/` and can be browsed with `tau2 view`.

The command uses the named model configuration; provider credentials must be configured.
This is a simulated airline-domain evaluation, not evidence of performance in live customer-service systems.

Evidence and transparency

Source record details

Source availability: Active
Sources checked: 11 October 2026
Original website title: sierra-research/tau2-bench · Official homepage title · retained in its source language