Harness-of-Harness lifts agents 52%

Harness-of-Harness (HoH) reported a 52.25% average relative gain, with a maximum 82.86% after three iterations across GameCraft-Bench, FrontierSWE, and ProgramBench, according to the authors’ arXiv paper. The framework repeatedly plans, codes, tests, and expands an existing agent harness, but these are preprint results rather than independent production evidence. For teams studying how harnesses shape coding-agent reliability, the practical takeaway is to benchmark the loop—not only the model—and budget for iteration cost before assuming a 52% lift.