Most work on LLM coding agents focuses on the policy model—the underlying language model that decides what to do. A new preprint on arXiv argues that the harness, the code that mediates the agent's interaction with tools, files, and execution environments, is just as important. The authors propose harness-policy optimization, a method that treats the harness as a learnable component and optimizes it together with the policy model.

Because the abstract is truncated, the exact training procedure and experimental results are not available from the source. What is clear is the central claim: that an agent's effectiveness is jointly determined by both components, and that improving them together—rather than fixing the harness and tuning only the policy—can lead to self-evolving coding agents.