Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails
A study across seven enterprise agent tasks found that evolving an agent harness with a weaker model, then having a stronger expert use it, often improves expert performance. However, training the weaker model on the expert's complete trajectories under the evolved harness caused performance regressions of 4 to 30 points on all seven tasks for Qwen3-Coder and Gemma 4, while the same procedure helped under the unevolved harness. The regression is attributed to imitation transferring knowledge but disrupting model-harness fit.
The research investigates how to combine automated harness evolution with lightweight fine-tuning for agentic tasks. It shows that a stronger expert can often use a harness evolved for a weaker model more effectively, suggesting expert supervision could close the gap. Yet, fine-tuning the weaker model on expert trajectories under the evolved harness backfires, causing performance drops across all tasks. The analysis indicates that imitation increases scaffold usage but disrupts the fit between the model and the harness, as the weaker model adopts the expert's planning strategy without the competence to execute it.
The key technical finding is that on-policy correction—where the weaker model is trained on its own trajectories under the evolved harness—may be necessary to preserve model-harness fit, whereas off-policy imitation from an expert under the same harness degrades performance. This suggests that harness evolution and model fine-tuning are coupled and must be co-optimized.
For enterprises deploying agentic AI, this implies that simply pairing a cheaper, fine-tuned smaller model with an evolved harness may not yield reliable performance. Instead, organizations may need to invest in co-evolution pipelines that maintain alignment between model behavior and harness design, potentially increasing development complexity but reducing inference costs.
The research highlights a cost-performance trade-off: smaller models with evolved harnesses can approach expert performance at lower cost, but naive fine-tuning can negate these gains. Businesses could save on inference costs if they adopt co-evolution techniques, but must avoid imitation-based training that breaks harness fit.
Future work may explore on-policy fine-tuning methods that incorporate harness feedback, or iterative co-evolution loops where harness and model are updated alternately. The observed regression could lead to new benchmarks for model-harness compatibility and more robust agent scaffolding tools.