Deploy an AI-built app to production only after treating the generated code like any other untrusted change. Confirm that requirements, tests, configuration, data migrations, security controls, monitoring, ownership, and rollback are ready. A successful local demo is not a production release criterion.
Build an immutable artifact through the documented pipeline. Pin dependencies, keep secrets in managed configuration, and separate development, staging, and production access. Run formatting, linting, types, tests, security checks, and the production build. Review authentication and authorization on the server, inspect logs for sensitive data, and test rate or resource limits where abuse matters. Rehearse migrations on representative data, verify backups, and define forward and rollback steps. Add health signals, error reporting, and an owner for alerts.
Release to a limited audience or percentage when the platform allows it. Test the main journey, failure paths, and external integrations in the deployed environment. Define measurable abort conditions before launch rather than deciding under pressure. Preserve the repository revision, artifact identifier, migration state, and approval record.
Keep the AI agent away from unrestricted production credentials unless a narrowly approved task requires access. Humans should authorize the release and assess live behavior. Once monitoring is stable and rollback is proven, expand traffic gradually. Production readiness is evidence that the system can fail safely, not evidence that code was generated quickly.
