Design an evaluation harness for a customer-support AI agent

A company's AI agent answers its customers over chat and takes actions through tools (refunds, plan changes, escalations). Every prompt, model, tool or policy change can break it. Design the harness that decides whether a change made the agent better or worse before it ships: scenarios, simulated customers, sandboxed tools, graders, several trials per scenario, comparison with the production baseline, release gating, and how production conversations feed new test cases. Work through cost and run time, and how you would tell a real 2-point regression from noise.