Harness-of-Harness: Multi-Day Autonomous Software Development with Continual Improvement
The framework reports a 52.25% average gain over standalone coding-agent harnesses after three iterations.
Harness-of-Harness wraps existing coding agents in repeated planning, coding, and testing loops, with emphasis on small verifiable increments and separate evaluation. The paper tests it across GameCraft-Bench, FrontierSWE, and ProgramBench using Codex with GPT-5.5, OpenCode with DeepSeek-V4-Pro, and Pi with MiniMax-M3. In a longer run of more than 70 iterations, the authors say it autonomously built a playable first-person-shooter game with story, mechanics, visuals, and audio. HF Daily Papers' note
Harness-of-Harness wraps existing coding agents in repeated planning, coding, and testing loops, with emphasis on small verifiable increments and separate evaluation. The paper tests it across GameCraft-Bench, FrontierSWE, and ProgramBench using Codex with GPT-5.5, OpenCode with DeepSeek-V4-Pro, and Pi with MiniMax-M3. In a longer run of more than 70 iterations, the authors say it autonomously built a playable first-person-shooter game with story, mechanics, visuals, and audio. HF Daily Papers' note
score 5