Recursive Harness Self-Improvement Boosts LLM Agent Performance

The study explores a novel approach called Recursive Harness Self-Improvement (RHI) to enhance the performance of large language model (LLM) agents. In the context of "model-harness co-evolution," harnesses are not just inference-time tools but also generate data that can influence future model training. RHI focuses on optimizing these user-constructed harnesses in a computationally lightweight manner. RHI works by representing the agent's operational loop as a prompt-level specification, which is then iteratively refined. This refinement process uses pairwise feedback based on the harness's own revision history, allowing for continuous improvement without extensive manual intervention or costly updates from model providers. Experiments across 30 synthetic machine learning research tasks demonstrated that even a few iterations of RHI substantially improved the performance of agents requiring low reasoning effort. These gains often surpassed the performance of agents operating at maximum reasoning effort, while simultaneously reducing inference costs by up to 60%. The primary mechanism for this improvement was identified as better task-specific context management and more effective information flow between agents, rather than simply longer reasoning chains.