Recursive Harness Self-Improvement Boosts Agent Performance

In the evolving landscape of AI, agent harnesses are not just tools for inference but also crucial components that can generate data to train future foundation models. This concept, known as model-harness co-evolution, suggests optimizing harnesses for both immediate agent performance and the quality of the data they produce. However, continually updating these harnesses can be resource-intensive. This research introduces Recursive Harness Self-Improvement (RHI) as a computationally lightweight solution. RHI focuses on optimizing user-constructed harnesses in a task-specific manner. It represents the harness as a prompt-level specification of the agent loop and iteratively refines it. This refinement process uses pairwise feedback derived from the harness's own revision history, allowing it to learn and improve over time with minimal computational overhead and few update iterations. Experiments across 30 synthetic machine learning research tasks demonstrated that RHI significantly boosts the performance of agents requiring low reasoning effort. It achieved results surpassing even maximum-reasoning-effort settings while simultaneously reducing inference costs by up to 60%. The primary mechanism for these gains was identified as improved task-specific context management and more effective information flow between agents, rather than simply extending reasoning traces.