Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails

New research shows that fine-tuning weaker agent models on complete expert trajectories under an evolved harness degrades performance, while an on-policy expert-correction pipeline — automated by a meta-level MLE agent — lets weaker models (Qwen3-Coder, Gemma…

Researchers propose combining automated agent-harness evolution with lightweight fine-tuning so that weaker agent models can target frontier-model agentic performance at lower cost. Across seven enterprise agent tasks using Qwen3-Coder and Gemma 4, naively training the weaker models on complete expert trajectories under an evolved harness regressed performance by 4-30 points on every task; both source reports agree on this figure. The regression is attributed to disrupted model-harness fit: under imitation, weaker models adopt expert planning without the competence to execute it. The authors instead introduce an on-policy expert-correction pipeline, automated by a meta-level MLE agent, in which an expert rewrites only the failing turn of the weaker model's rollout — combining harness evolution with weight adaptation while preserving model-harness fit and the model's planning style. The method targets economical co-evolution for domain-specific enterprise agents, aiming for frontier-model agentic performance with smaller models.

• Paper: 'Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails' (listed on Hugging Face daily papers 2026-09-07 and arXiv cs.AI/cs.LG/cs.CL 2026-09-08).
• Evaluated across seven enterprise agent tasks.
• Weaker models studied: Qwen3-Coder and Gemma 4.
• Naive imitation fine-tuning on complete expert trajectories under an evolved harness regressed performance by 4-30 points on all seven tasks (consistent across both sources).
• Imitation disrupts model-harness fit: weaker models adopt expert planning without the competence to execute it.
• Proposed on-policy expert-correction pipeline is automated by a meta-level MLE agent.
• Correction rewrites only the failing turn of the weaker model's rollout, preserving model-harness fit and the model's planning style.
• Goal: economical co-evolution — smaller models with evolved harnesses targeting frontier-model agentic performance at lower cost.