Imitation can break the agent harness it was meant to improve
A smaller model trained on an expert’s full trajectories regressed by 4 to 30 points across seven enterprise tasks, despite the same recipe helping before harness optimization. The problem was not simply weaker capability: imitation changed the model’s planning style, breaking its fit with the system prompt, tools, and execution scaffolding tuned around its native behavior. Salesforce’s fix uses expert feedback only on the failing turn, preserving that fit while adding targeted corrections.