Co-Harness: Jointly Optimizing Harness and Weights for LLM Agents

A novel framework named Co-Harness enhances both agent harness and model parameters post-training for automated AI research. Traditional methods train models using a static harness—comprising prompts, tools, skills, middleware, and memory—while neglecting the optimization of the data-generating process. Co-Harness employs a dual approach, alternating between optimizing the harness and the model. A HarnessCritic, based on LLM, evaluates unsuccessful trajectories, pinpoints failure modes at the harness level, and suggests validated local improvements. Subsequently, the model is fine-tuned using high-quality trajectories derived from the enhanced harness, effectively refining behaviors. This strategy resolves the disconnect between model updates and the unchanging scaffolding that influences trajectory quality.