Tau-Tau evaluates six agent-building configurations
The strongest tested builder achieved 23.9% versus experts at 82.2% across 53 tasks in four domains.
Tau-Tau-Bench evaluates whether coding agents can build agents that meet a specification. Its 53 tasks span four domains. The reported strongest builder result is 23.9%, compared with 82.2% for expert-built agents under the evaluation.
Section 4 tests six configurations, including different model tiers and coding harnesses. It covers GPT-5.6 Sol and GPT-5.6 Terra through Codex, Claude Opus 5 and Sonnet 5 through Claude Code, and Kimi K3 through Kimi Code and OpenCode. The original article incorrectly said model/toolchain variation had not been tested.
These comparisons establish a gap in the tested agent-building tasks. They do not establish that every generated agent is unsuitable for production. Inspect held-out acceptance conditions, runtime behavior and the relevant task family before applying the aggregate result.
Editorial correction (2026-10-13 UTC): Restored the six tested model/harness configurations.