The paper studies agent harnesses (system prompt, tools, execution hooks, context management) together with model fine-tuning, finding that naively imitating an expert’s full trajectories under an evolved harness backfires — performance regresses 4-30 points across seven enterprise agent tasks and two model families, because imitation disrupts the model’s fit to its own harness. The authors instead build an on-policy expert-correction pipeline that localizes only the failing turn in the weaker model’s own rollout and has the expert rewrite just that turn, preserving the model’s planning style while combining the benefits of harness evolution and model adaptation.
