The paper introduces Harness-Zero, a method for transferring behaviors learned with an optimized external agent harness into a model’s own weights so it can be deployed with a simpler harness. An intermediate agent converts guidance from the specialized harness into training demonstrations in the target harness’s action space, enabling distillation despite differing action spaces. Experiments across multiple domains show recovered harness-induced behaviors and substantially higher task success even after the specialized harness is removed at deployment.
