Running AI-generated backend code safely requires more than a sandbox. The post describes a verification loop that combines typed contracts (catching shape errors at compile time), real infrastructure (Postgres via Docker instead of mocks), per-PR preview environments with realistic data, and an MCP server that lets the agent inspect live schema, call its own endpoints, and read traces. Together these replace the agent's assumptions with observable facts, so code earns its way to production by surviving real checks rather than just passing mocks. Business logic correctness still requires human review, but reviewers see evidence of behavior rather than raw generated code.
Table of contents
Catching mistakes before it runsRunning against real dependenciesLetting the agent check its own workWhat a person still reviews24 Impressions