Salesforce Discovers a Better Way for Co-Evolution of Agents and Harnesses (9-minute read)

Salesforce researchers conducted a study combining Agent Harness (system prompts, toolsets, execution hooks, etc.) with lightweight fine-tuning to improve the performance of smaller models. In existing approaches, training weaker models on expert model trajectories under evolved harnesses via Imitation Learning resulted in performance degradation.

Causes of Performance Degradation

A performance drop of 4–30 points was observed across all 7 enterprise tasks for Qwen3-Coder and Gemma 4 models. This occurs because weaker models adopt the expert's planning strategy but lack the capability to execute it, leading to a mismatch with harnesses evolved to fit their original planning style.

On-Policy Expert-Correction Pipeline

To address this issue, the researchers proposed an automated pipeline leveraging a Meta-level MLE agent. This approach identifies failed turns in the weaker model's own rollouts and requests the expert model to rewrite only those specific turns. This technique preserves the model's original planning style while combining the benefits of harness evolution and model adaptation, enabling cost-effective co-evolution for domain-specific tasks.