Agent harnesses — the external systems that mediate how a model interacts with its environment — can meaningfully boost agent performance, but those gains are typically tied to the harness present at deployment. Because the best harness varies by domain, instance, and model, a general-purpose agent must either accept a suboptimal shared harness or route among a growing set of specialized ones.
To address this, researchers introduce agent harness distillation: using a domain- or instance-optimized harness as training-time guidance and transferring the behaviors it induces into model weights, so its benefits survive under a single fixed target harness. The core challenge is that the optimized harness and the target harness differ in action space and available information, meaning guidance from the former cannot be used directly as supervision for the latter.
The proposed method, Harness-Zero, enables harness distillation through an "agent-as-harness" approach. Guided by the optimized harness, a harnessing agent corrects student responses before execution in the target harness's action space, converting harness guidance into training demonstrations. Fine-tuning on the resulting trajectories internalizes harness-induced behavior into the model, allowing the specialized harness to be removed at deployment.
Experiments spanning knowledge work, tool use, and science domains report three main findings. First, for frontier LLMs using the same evolved harness, agent-as-harness outperforms code-as-harness. Second, with the specialized harness removed at deployment, Harness-Zero raises the base model's macro-average task success from 23.3% to 44.3% — even exceeding the 41.7% achieved when that harness remains attached. Third, Harness-Zero recovers harness-induced behaviors absent from the base model, with 82.3% average recovery across 28 patterns in the three domains.