Cursor to production: turn agent-assisted code into a release with owners

Local agent success hides missing review: secrets pasted into prompts that later leak into commits, flaky tests skipped to “just ship,” and env drift between a developer machine and staging. The first production tickets are often unauthorized writes, broken migrations applied once by hand, and no one knowing which PR introduced the regression because agent commits lacked clear ownership notes.

It means the remote repository is the source of truth (not a single Cursor workspace), secrets live in a vault or CI store never echoed into agent context unnecessarily, every merge to the release branch passes automated checks you defined, and rollback is a practiced promote-previous—not a frantic local revert. “It works in my Cursor session” is not the bar.

Harden when the architecture and data model still match the product. Add review gates, expand tests around agent-touched modules, and quarantine high-risk paths (auth, billing, migrations) for human-led diffs. Rewrite when the repo is a pile of unreviewed agent patches with no module boundaries. Prefer narrow ownership seams over a panic rewrite of everything Cursor touched.

First value is one real user completing the core journey on a production (or production-like) environment provisioned from CI, with secrets injected outside the IDE, test evidence attached to the release, and a rollback path an on-call can execute without opening Cursor. Instrument that path before optimizing agent prompts for secondary features.