τ-bench

Hold

Techniques

Sierra’s earlier benchmark for measuring finished agents on user interaction and tool use.

Why it's here

Placed in Hold: 1 article(s) of evidence from 1 source(s), led by research-stage coverage, with 0 in the last 30 days. Confidence 24%. Low accumulated evidence, so it defaults conservatively pending more signal.

Evidence (1)

7The New Stack·9/9/2026research
Claude Leads New Benchmark for Agents That Build Agents
Sierra’s new Hyper-τ-bench evaluates how well AI developer agents can build customer service agents from simulated business materials. In early results, Claude Opus 5 in Claude Code scored highest at 23.9%, while all six autonomous configurations stayed below 25%, far behind the 82.2% oracle reference.