Harness Policy Distillation: aligning agent policies with the semantics of the execution environment

What was downloaded is the arXiv abstract page for the paper, not the full text. This analysis is based solely on the abstract and the metadata on that page. I cannot describe implementation details, ablations or any other experiments the paper may contain.

The paper is titled “Harness Policy Distillation: Aligning Agent Policies with Runtime Semantics.” Its authors are Junze Zhu, Weihao Chen, Zhiyang Zhou, Zhen Wu and Xinyu Dai. It was submitted to arXiv on October 14, 2026, in the machine learning category (cs.LG), with cs.AI as a secondary category. It runs to 23 pages and 9 figures.

The authors' starting point is a practical one. When a general-purpose model is deployed as an agent, it runs inside a product-specific execution “harness.” That harness determines which actions are executable, how state advances and when the agent loop can terminate. According to the authors, standard trajectory supervision only shows which action was taken. The execution semantics behind that action remain implicit.

To address this, they propose Harness Policy Distillation (HPD). The idea is to turn those control semantics into an explicit training target, alongside the actions, in order to align the harness policy. According to the abstract, the method involves these steps:

• Reference trajectories are replayed through the harness to recover the execution semantics that ordinary trajectory data leaves implicit.

• Those semantics are represented as “control documents” at each model-call boundary.

• These documents serve as privileged information for a “teacher” model to generate actions. In addition, harness verification ensures that the resulting targets are executable under the execution contract.

• A shared model is trained jointly on the semantics recovered from the privileged control-state information and on the native history. This internalizes those semantics into the model's policy.

At deployment, the trained model uses the harness's original interface. It does not need the control document or any additional supplement. This is the core of the approach: the privileged information is used only during training.

The only quantitative result given in the abstract is this. On TRAJECT-Bench with the Pi harness, Qwen3-8B trained with HPD achieves 22.98% autonomous task success and 49.23% runtime compliance. The unadapted model scores 3.09% and 19.90%. The authors add that harness-verified teacher actions substantially improve the deployed native policy. According to them, joint control-state supervision brings further gains in task completion and compliance.

The relative improvement is large, but the absolute level remains modest: fewer than a quarter of tasks are completed autonomously. Nor can I assess from the abstract alone how HPD compares with other adaptation techniques, or whether the results hold for other harnesses or models. That information would have to be sought in the full text.

The abstract presents HPD as a way of turning execution semantics into dense supervision for learning harness-compatible actions at every decision point. The conclusion that can be drawn without reading the full paper is limited. Part of an agent's performance could be improved by aligning the model with the rules of the environment it runs in, without changing the base model.