← All notes

Summary

τ-bench (tau-bench) is a benchmark for tool-using language agents in dynamic conversations with a simulated user, using domain-specific APIs and policy guidelines. The authors evaluate success by comparing the final database state with an annotated goal state, and they propose a pass^k metric for reliability across multiple trials. They report that even state-of-the-art function-calling agents such as gpt-4o succeed on less than 50% of the tasks, and that pass^8 is under 25% in retail. [S1]

Claims

• id: C1 claim: 'τ-bench emulates dynamic conversations between a simulated user and a tool-using agent.' source: S1 evidence: "dynamic conversations between a user (simulated by language models)"
• id: C2 claim: 'They report that even gpt-4o-style function calling agents succeed on less than 50% of the tasks.' source: S1 evidence: "even state-of-the-art function calling agents (like gpt-4o) succeed on <50% of the tasks"
• id: C3 claim: 'Code and data are pointed to a public GitHub repository.' source: S1 evidence: "Code and data: https://github.com/sierra-research/tau-bench"

Verification

• Code available: Paper: https://github.com/sierra-research/tau-bench, which returned HTTP 200 on 2026-11-11. Not run.
• Reproduced: not reproduced (2026-11-11). Code not run.
• Related work: Live notes 2305-15334-gorilla and 2402-01030-codeact. This is the paper-verification note for τ-bench itself.
• Limits: The headline reliability numbers are from the v1 paper (17 Jun 2024) and a specific retail split. Later models or prompts may change pass^k.
• License: CC BY 4.0 on the abs page (2026-11-11; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 17 Jun 2024 (v1). Sources: [S1] https://arxiv.org/abs/2406.12045, v1 (17 Jun 2024). checked_on: 2026-11-11.

When to apply

Use when asking whether a tool-using agent follows domain rules consistently across trials, not only whether it can call an API once.

Facts

• id: F1 claim: 'τ-bench tests tool-using agents in multi-turn conversations with a simulated user and domain policy guidelines.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F2 claim: 'The authors propose a pass^k metric for reliability over multiple trials.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F3 claim: 'They report success under 50% for state-of-the-art function-calling agents, and pass^8 under 25% in retail.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0

Sources

• [S1] arXiv 2406.12045, https://arxiv.org/abs/2406.12045. License: CC BY 4.0. v1 (17 Jun 2024). Accessed 2026-11-11.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.

@misc{iil_note_q4vw0dyz,
  author = {{'iinventedlight (AI-assisted)'}},
  title = {{τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-11},
  url = {https://iinventedlight.com/library/notes/2406-12045-tau-bench/},
  note = {Permanent ID iil:note:q4vw0dyz. License: CC-BY-4.0.},
}

Markdown

Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.

For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/2406-12045-tau-bench.md.