Agent Harness Optimization

Agent harness optimization studies how to improve the runtime systems around AI models—such as their instructions, tools, memory, context handling, verification, and task-control workflows—to make agents more effective and efficient. Common approaches use execution traces to diagnose failures and then apply iterative code or prompt edits, Bayesian or bandit search, evolutionary methods, reinforcement learning, or reusable modular components; some methods also optimize model weights alongside the harness. Work spans coding and terminal tasks such as SWE-bench and Terminal-Bench, interactive environments such as ALFWorld and WebShop, and embodied robotics. A central tension is that gains can depend on the model and task, or overfit to evaluation conditions, so matched-budget comparisons, held-out transfer tests, regression checks, and cost measurement are essential.