Shrinking an AI agent usually makes it forget how to use its own tools - a new paper claims a fix.
Today's AI agents run inside a harness, the surrounding software that manages their context, calls their tools, and feeds back results. When a big model gets distilled into a cheaper one, the harness stays in place, but standard distillation still forces the smaller 'student' model to copy everything the teacher model did, including things the harness already tells it. The new method, called Harness-Aware Distillation, instead trains the student only on what the teacher adds beyond what the harness already provides. It works by comparing the same teacher's actions with and without harness information, scoring that comparison only after the student's own reasoning, then filtering out any comparisons that contradict the harness's own records. Tested across several long-horizon agent benchmarks, the method beat standard on-policy distillation using the same fixed harness, got stuck in fewer repetitive loops, and recovered from mistakes more often.
That last part is the real finding. Most distillation methods need task rewards or labeled examples of success to steer a smaller model, which is expensive and often unavailable outside a lab. This method needs none of that, which matters for anyone trying to run cheap agents in production rather than benchmarks, where labeled outcomes are rare and harnesses differ from one deployment to the next.
It is still a preprint, not a shipped product, and the benchmarks are not the messy production harnesses most companies actually run. Whether the gains survive contact with a real customer-support bot or coding agent is the next question, not this one.