Your Policy Is Secretly a Harness: The Decision Geometry of Agent Interfaces

Your Policy Is Secretly a Harness. The Decision Geometry of Agent Interfaces. By Jian Mu and two others, from the Hong Kong University of Science and Technology, Guangzhou and the Hong Kong University of Science and Technology. Published on the twenty-seventh of October.

A language model agent never meets a task directly. It is post-trained and then deployed through a harness: the prompts, the tools and the control logic that present the task to the model and carry out what the model outputs. Harnesses get revised all the time, and a new one is usually judged by how an already trained model performs inside it. That habit assumes that whatever the model learned under one harness carries over to the next. The authors ask what a policy actually learns from its harness, and what survives when the harness changes.

Their answer, worked out under KL-regularized control, is that a trained policy carries the preferences of its harness. The log odds of an optimal policy recover how strongly the harness favors each action over the others. From there the deployed return splits into two terms. One is the ceiling of the harness, the best return it can offer at all. The other is the mismatch of the policy, the return lost to its own suboptimal decisions. Deployed return is the ceiling minus the mismatch. Training only changes the mismatch. Changing the harness moves both. That is why comparing harnesses with a fixed model is muddled: it mixes what each harness offers with how well that particular model happens to use it.

In MiniWoB++, post-training makes a model's preferences specific to the harness it was trained under. The sharpest demonstration is almost cosmetic. Reversing the order of the observation fields leaves every optimal decision exactly as it was, yet a trained model's success falls from eighty-five percent to nineteen point two percent. Nothing about the task changed. Only the presentation the policy had grown used to.

The decomposition also predicts the other direction. Some new harnesses are rejected by a fixed-model comparison even though they raise the ceiling. Every one of those, after further training under it, outperforms the original agent. And the ceiling and mismatch, measured beforehand, predict how large those gains will be. For coding-agent harnesses on Terminal-Bench Lite, the authors estimate ceiling minus mismatch using a stronger model as a stand-in for the optimum. That estimate reproduces the accuracy ranking of the harnesses, while neither term on its own does.

The theoretical result is stated under KL-regularized control, and its statements concern an optimal policy. The empirical support comes from two settings: MiniWoB++, and coding-agent harnesses on Terminal-Bench Lite. In the second, the optimum is not available directly, so a stronger model stands in for it, and the ranking result rests on that substitution. What the work offers is an accounting of return into ceiling and mismatch, and a warning about fixed-model comparisons, rather than a recipe for designing any particular harness.

The title's inversion is the point. One might think of a harness as packaging around a policy, something to swap freely. On this account, after training the policy is partly a record of the harness that shaped it. Evaluating a new harness then means asking two separate questions: how high its ceiling sits, and how far a given model still is from reaching it.