There's an awkward moment in every agentic data project. The agent works. It writes decent SQL, it reasons about schema, it proposes a migration that looks right. And then somebody asks the question nobody wants to answer: what happens when it's wrong against production?The usual answers are all bad.