Whenever you're ready, walk me through how you'd approach this.
How do you evaluate the reliability and hallucination rate of an autonomous agent before deploying it into a production environment?
Design a practical, reproducible evaluation framework rather than relying on anecdotal testing. Explain how you would construct representative evaluation data, define hallucinations and task failures, test tool use and refusal behavior, measure uncertainty, and report confidence intervals. Address adversarial cases, distribution shift, nondeterminism across repeated runs, evaluator bias, and deployment gates. Provide production-quality Python that calculates the core metrics from agent-run records.