Harness-of-Harness Framework Delivers 52% Average Gain on Multi-Day Autonomous Software Dev
Summary
The Harness-of-Harness (HoH) paper wraps existing coding-agent harnesses in an iterative planning-coding-testing loop with versioned project histories, splitting implementation testing from independent evaluation. Across GameCraft-Bench, FrontierSWE and ProgramBench, HoH reports an average relative gain of 52.25% and a max of 82.86% after three iterations, and used 70+ iterations to autonomously ship a playable first-person shooter with story, mechanics and audio.
Originally reported by huggingface.co
Read the original article →Original headline: Harness-of-Harness Framework Delivers 52% Average Gain on Multi-Day Autonomous Software Dev