BuilderTool use · benchmark
τ-bench
Evaluate customer-service agents on tool use, policy compliance and user interaction.
What it measures
Pass^1 and reliability
Current τ³ text, knowledge and voice tracks are separate from legacy τ².
Public evaluation code. Check upstream code and dataset terms.
Open source / dataset · Run instructions · Official results
Reported ranking tracks
- τ³-bench · banking text · 3 entries · captured 2026-10-07T07:01:28Z
- τ³-bench · voice · 3 entries · captured 2026-10-07T07:01:28Z
- τ²-bench · text · 3 entries · captured 2026-10-07T07:01:28Z