// benchmark explainer
τ²-Bench (tau2), explained
By the TrueIQ editorial team at Ahzy Capital · Last reviewed Oct 15, 2026 · How we work
τ²-Bench tests whether an AI agent can handle a customer service conversation correctly. It has to follow a written policy, use tools such as databases, and deal with a simulated customer.
| Made by | Sierra (successor to τ-bench) |
|---|---|
| Domains | Airline, retail and telecom scenarios |
| Metric | Task success; pass^k checks reliability over repeated tries |
| Code | github.com/sierra-research/tau2-bench |
What it measures
It tests reliable tool use and rule following over a multiturn conversation, which matters a lot for real business agents.
Where it falls short
- The simulated user is itself an AI model, which adds noise.
- Only a few domains are covered.
- Single run scores hide inconsistency. Look at pass^k when it's available.
How to read a score
Prefer pass^k (success on all k tries) when judging reliability. A model that succeeds once in four tries isn't ready for customers.
Where to see current scores
We don't republish scores. Leaderboards change weekly, and numbers depend on how the test was run. Check these instead:
- Artificial Analysis: τ²-Bench (tau2) results (independently run)
- τ²-Bench repository