tau2-bench (τ²-Bench)

A dated reading of what is claimed, reported, and independently verified in the current evidence.

Current reading

The readiness gap, in one scan

AI-assisted assembly · derived results

Evidence strength Strong

Decision

Human editorial judgment · 2026-06-28

Adopt with guardrails

Why: It is the de-facto industry standard for conversational tool-agent eval and its pass^k/dual-control framing maps onto our shared-world agent risks, but headline numbers are domain-, simulator- and harness-sensitive and must not be quoted bare.
Next: Use τ²-Bench's dual-control + pass^k methodology as the template for our customer-facing agent reliability gates; report pass^k (not pass^1), segment by domain difficulty, and pin a fixed user-simulator + harness version rather than citing a single leaderboard figure.

Constraints

No named blocker is present in the current public projection.

Evidence summary

AI-assisted assembly

Counts and dates only. Raw signals, private excerpts, trust records, and internal corpus material are not published here.

Publication record

Initial public reading

This is the initial public reading. No earlier readiness change is recorded.