The Training Trace Had to Match the Harness That Made It

Terminal agents learn inside a harness that formats prompts, binds tools and recovers from errors, making a trajectory's value conditional on the runtime that produced it. CoTrace alternates harness search and policy training while matching supervised rollouts to the adopted harness and using fresh interactions for reinforcement learning. On the Tmax promotion split, the authors report Qwen3.5-9B advancing from 78 to 88 solved tasks under supervised fine-tuning and to 90 with online reinforcement learning. A compact harness-matched corpus outperformed much larger pools from sibling harnesses, and out-of-distribution tests showed failures when training and evaluation scaffolds diverged. The result treats harness provenance as training data, not incidental metadata.