Doc2Harness Benchmark: Top Models Pass Only 53% on Coding Agent Harness Tasks

ICT-CAS, Tsinghua and Peking University introduced Doc2Harness, the first deterministic benchmark for long-horizon harness engineering, testing 19 frontier models across 328 tasks. Even the best performer, GPT-6 Astra, passed only 53%, with nearly half of all tasks still failing.

2026-10-15 ~ 2026-10-16 · 2 related posts

• Doc2Harness benchmarks LLMs on harness engineering: GPT-6 Astra tops at 53.05%, still failing half — Yuyao Ge · 2026-10-15
• Doc2Harness: New benchmark finds best coding agent passes only 53% across 329 cases — _akhaliq · 2026-10-16