Design an agent harness and its evaluation system

Design the harness that runs an AI agent (the loop that calls the model, runs its tools and manages its context) together with the system that evaluates it. Cover how the agent executes tasks, how runs are recorded, and how you measure and compare agent versions.