CoTrace: Data Recipes for Training Terminal Agents with Harness-Model Co-Evolution
arXiv cs.CL · 8 Oct 2026, 03:06 AEST
CoTrace introduces a framework for improving terminal-agent capabilities through harness-model co-evolution. This approach enhances model training by explicitly governing trajectory routing and ensuring compatibility between training and evaluation runtimes.
Key facts
- Terminal-agent capability depends jointly on model weights and the runtime harness that formats prompts, binds tools, and handles error recovery.
- Existing harness-model co-evolution approaches improve both components, yet often treat trajectories produced during harness search as an undifferentiated replay buffer.
- Under CoTrace, recurring execution failures guide harness synthesis, while policy training is strictly conditioned on verified rollouts matched to the adopted runtime for supervised fine-tuning (SFT) or fresh online interactions for reinforcement learning (RL).
- On the Tmax promotion split, CoTrace advances Qwen3.5-9B from 78 to 88 solved tasks under supervised fine-tuning while an online reinforcement variant reaches 90.
- Evaluations on Terminal-Bench 2.1 and SWE-bench Lite show that out-of-distribution transfer depends fundamentally on harness compatibility.
Read aloud
CoTrace introduces a framework for improving terminal-agent capabilities through harness-model co-evolution. This approach enhances model training by explicitly governing trajectory routing and ensuring compatibility between training and evaluation runtimes. Terminal-agent capability depends jointly on model weights and the runtime harness that formats prompts, binds tools, and handles error recovery. Existing harness-model co-evolution approaches improve both components, yet often treat trajectories produced during harness search as an undifferentiated replay buffer. Under CoTrace, recurring execution failures guide harness synthesis, while policy training is strictly conditioned on verified rollouts matched to the adopted runtime for supervised fine-tuning SFT or fresh online interactions for reinforcement learning RL. Evaluations on Terminal-Bench two point one and swe bench Lite show that out-of-distribution transfer depends fundamentally on harness compatibility.
Sources in this cluster
- CoTrace: Data Recipes for Training Terminal Agents with Harness-Model Co-Evolution
arXiv cs.CL · 8 Oct 2026, 03:06 AEST