Autonomous software development has long been a goal for AI researchers, but most coding agents struggle to turn high-level requirements into fully functional software without human help. A new paper introduces Harness-of-Harness (HoH), a framework designed to let coding agents continually improve software over multiple days of autonomous work.
HoH works by wrapping existing coding-agent harnesses and organizing their executions into iterative planning-coding-testing loops. The framework balances repair with capability growth, breaks development into small, verifiable increments, and separates implementation-time testing from independent evaluation. Instead of prescribing agent workflows, HoH constrains verifiable outputs, progressively exposes deliverables, role-specific tools, and skills, and encourages reuse over recreation while maintaining versioned project histories.
In experiments, HoH was tested on three benchmarks—GameCraft-Bench, FrontierSWE, and ProgramBench—using three harness-model pairs: Codex with GPT-5.5, OpenCode with DeepSeek-V4-Pro, and Pi with MiniMax-M3. Across these combinations, HoH consistently outperformed the standalone harnesses, achieving an average relative gain of 52.25% and a maximum gain of 82.86% after three iterations.
In a multi-day deployment with more than 70 iterations, HoH autonomously developed a complete first-person-shooter game, featuring a coherent storyline, fully implemented core mechanics, a human-playable experience, polished visuals, and integrated audio. This demonstrates the framework's ability to handle complex, long-horizon projects.
The code is available on GitHub and the project page provides further details.