Successful AI-agent demonstrations do not establish operational reliability. A reproducible evaluation framework for workflow constraints, failure handling, authorization, and evidence-based deployment decisions.