Using an LLM to iteratively improve an agent's harness components based on evaluation feedback within a fixed computational budget.