How do you stop AI coding assistants from "over-optimizing" your database models?

We ran into a near disaster last week when an agent was tasked with a routine API validation fix, but took it upon itself to "clean up" the database schemas and generated unrequested migrations.

It completely wrecked a couple of foreign key relationships because it thought it knew a better architectural pattern. Ever since then, our team had to lock down write permissions for any automated tool.

For those of you doing heavy backend work with AI, what kind of strict guardrails or sandbox environments do you use to keep the assistant away from your data layer?