A new paper introduces CoTrace, a framework for training terminal (command-line) coding agents that explicitly matches training data to the runtime harness used at deployment, rather than treating all generated trajectories equally. The method uses an alternating co-evolution process in which recurring execution failures during agent operation guide the synthesis of new harness code, while policy training is restricted to verified rollouts matched to the adopted runtime, for use in either supervised fine-tuning or reinforcement learning. In evaluations, smaller harness-matched training sets produced higher task-solve rates than larger mixed-source datasets while using considerably less computational resources.
