You deploy merged code straight to production. No staging environment? That is how you break customers.?

The full executive answer

There is a preview layer — it is just per-change rather than a shared bottleneck. Every pull request gets its own complete, isolated preview deployment on Vercel: a full running instance of the application at its own URL, built from that exact change. That is arguably stronger than a classic shared staging environment, which the industry has been retiring for well-known reasons: shared staging drifts from production, queues teams behind each other, and produces 'worked on staging' incidents. Preview-per-change plus production is the model Vercel-native shops run, and it means every change is inspectable as a running system before merge.

What guards the production doorstep, given no human staging sign-off: the five automated gates — lint, strict types generated from the real database schema, the AI regression harness, the full test suite, and a production build — followed post-merge by a browser-level smoke test against a true production build verifying the security-critical routes. Then the blast-radius controls: deployments are immutable, so rollback is re-pointing to the previous build in minutes; database migrations are written idempotent — safe to re-run — which removes the classic deploy-rollback schema trap; and per the DORA research, this small-batch model empirically lowers change-failure rate versus big staged releases, because each deploy carries one change, not forty.

The honest gap in the model at current scale: no canary or progressive rollout — every deploy goes to 100% of traffic at once. With today's traffic that is proportionate; traffic-splitting rollout is the natural hardening step as production volume grows, and the platform's deployment layer supports it when we turn it on.