Vibe coding needs a senior engineer on the hook

The demo everybody has seen by now: a prompt, ninety seconds, a working app. It is real. We use it every day, and it is why a project that used to take a quarter now takes days.

The part the demo leaves out is who answers when it breaks at 2 a.m. in month four.

Speed is the easy half

Generated code is fast to produce and fast to accept. Both halves of that are dangerous. A model will happily write an auth check that looks right, a migration that drops a column it should rename, or a retry loop with no ceiling. Each one reviews cleanly if you are skimming for style instead of reading for consequence.

The failure is not that the model is bad at code. It is that review effort has to scale with output, and it does not scale on its own.

What accountability looks like in practice

On our projects one senior engineer owns the branch. Not “reviews the PR” — owns it, the way you own code you typed yourself. Concretely:

• Every generated change gets read against the failure it could cause, not against a style guide.
• Tests are written for behavior we decided on, before the implementation is generated. A model asked to write its own tests will write tests that pass.
• Anything touching money, auth, or data deletion is written by hand or rewritten by hand. The time saved there was never the bottleneck.
• The engineer who owns it can explain any line in it without opening the transcript.

The hardening pass is not optional

Working software and shippable software are separated by a specific, boring list: error paths, timeouts, idempotency, observability, the migration rollback, the thing that happens when the third-party API returns a 500 for six minutes. AI gets you to working in days. Shippable still costs what it always cost, and pretending otherwise is how teams end up with a prototype in production.

Budget for it, out loud, in the estimate. It is the difference between a fast project and a fast project you can leave running.

This is how we run our vibe coding and AI-native delivery work — one senior engineer on the hook per branch, hardening priced into the estimate rather than discovered in month four.