τ²-bench (tau-bench)

Simulated customer-service benchmark where an agent follows domain policies and uses tools while a simulated user also acts.

Reference by Sierra

IndexedMIT★ 2,310

About

Whether a conversational agent completes tasks correctly and follows policy in airline, retail and telecom domains, over text or voice.

Use cases: llm-evaluation, customer-support

Reviews

No reviews yet. Installed it? Reviews come from verified installers only. How reviews work

About the benchmark

Benchmark site: taubench.com/
Current version: v2.0.0 (adds six languages and voice)github.com
Maintainer: Sierrataubench.com
Launched: 2025-06arxiv.org
Paper: arxiv.org/abs/2506.07982
Leaderboard: taubench.com/

We link to the official leaderboard instead of copying scores. Benchmarks get gamed and saturate; we never use them to rank listings.

Signals

Stars: 2,310
Forks: 606
Open issues: 132
Last commit: Oct 24, 2026 (4 days ago)
Last release: Oct 20, 2026 (8 days ago)

From GitHub (fetched Oct 24, 2026).

Details

Version: v2.0.0 (adds six languages and voice)
License: MIT
Repository: github.com/sierra-research/tau2-bench
Website: taubench.com/
Docs: taubench.com/
First seen: Oct 24, 2026
Last checked: Oct 24, 2026
Source: Specification Paper github-graphql

More from Sierra

0.450score
Sierra
Sierra
Conversational AI platform where businesses build customer service agents that run across chat, SMS, email and voice.
AgentCFreshness: C

Cite this page

τ²-bench (tau-bench) benchmark. Agent Mart. https://agentmart.dev/reference/tau2-bench (last updated 2026-10-24)