Harmony Harness is a framework for multi‑objective, compositional evolution of agent harnesses. It formulates harness evolution as a multi‑objective optimization problem and applies decomposition‑based evolutionary search to preserve diverse solutions along the Pareto frontier, removing the need for hand‑crafted structural diversity metrics. In parallel, factorized evolution operators separate general‑purpose mechanisms from benchmark‑specific adaptations, which are dynamically composed at inference time. Experiments on SecRepoBench and BaxBench show that Harmony Harness consistently outperforms strong baselines by more than 10%. The results demonstrate that explicitly modeling objective trade‑offs and harness structure yields substantially stronger and more transferable agent evolution.
Review: This study advances self‑evolving agents by combining multi‑objective decomposition with factorized operators, offering a scalable and adaptable approach for future autonomous systems.