Salesforce researchers, led by Zhou Yu, published a preprint arguing that a common recipe for cheapening AI agents, imitation training, can actively damage them. The paper, posted to arXiv on September 8 and not yet peer reviewed, examines “agent harnesses”: the system prompts, tools, and scaffolding wrapped around a model that shape how it performs a task.
The team first optimized a harness using a weaker model, then had a stronger model use that same harness. The stronger model did better with it, which suggested the gap would close if the weaker model were simply fine-tuned on every full trajectory the stronger one produced.
Instead, the paper reports, that training regressed performance by 4 to 30 points across Qwen3-Coder and Gemma 4 on all seven enterprise tasks tested. The authors say imitation pushed the weaker model to adopt a planning style it could not execute, breaking its fit with a harness built around its own native approach. Their proposed fix repairs just the step where the weaker model’s output goes wrong, leaving the rest of its strategy intact.
Reported from a preprint published on arXiv on September 8, 2026, which has not undergone peer review.
