Training coding agents typically updates the policy while leaving the harness—the surrounding interaction and tool-use framework—fixed. A new arXiv paper (2610.28069) points out that this common practice overlooks what training trajectories show: flawed repair strategies and interaction problems that are not static.
The abstract notes that these issues vary across tasks and change as the policy itself learns. That suggests a fixed harness may become misaligned with the evolving policy, leading to inefficiencies or failures that a model-only update cannot address.
The paper's title, "Training the Whole Agent, Not Just the Model," signals a call to include the harness in the training loop. However, the abstract is truncated, so the specific method proposed is not described in the available text.