Harness optimization lifts legal agent benchmark from 67% to 86%

Adaption AI shows that optimizing only the agent harness raised pass rates on the Harvey legal benchmark from 67.1% to 85.9% with Qwen3.6-27B. Sara Hooker cautions that harnesses can only unlock existing model capabilities, arguing the whole stack should be treated as optimizable.

2026-09-09 ~ 2026-09-10 · 3 related posts

• Harness tuning lifts Qwen3.6-27B legal agent pass rate from 67% to 85%, study finds — sarahookr · 2026-09-09
• Harness optimization lifts Harvey legal agent benchmark pass rate from 67.1% to 85.9% — sarahookr · 2026-09-10
• Sara Hooker: optimize the whole stack, not just the model or the harness — alienelf · 2026-09-10