Wrapping existing coding harnesses in a planning-coding-testing outer loop gains 52% average across three model pairs

arXiv·high signalXBlueskyLinkedInCopy link

Harness-of-Harness runs on top of existing coding-agent harnesses, organizing their executions into iterative planning-coding-testing loops that balance repair against capability growth, scope work into small verifiable increments, separate implementation-time testing from independent evaluation, and constrain outputs rather than prescribing workflows. Across GameCraft-Bench, FrontierSWE and ProgramBench with Codex+GPT-5.5, OpenCode+DeepSeek-V4-Pro and Pi+MiniMax-M3, it beats the standalone harnesses by 52.25% on average and 82.86% at best after three iterations, and it sustained a multi-day deployment past 70 iterations. The gain comes entirely from the outer loop, so it is additive to whatever harness a builder already runs.