AI Coding Weekly

Rauch sketches a three-step loop for agents shipping code

Feature-gated code, verification gates on push, then the agent watches the metrics.

Guillermo Rauch posted a three-step description of how agents will ship all code. The agent writes feature-gated code. Verification gates run on push, which he describes as proof checking and security audits. Then the agent monitors metrics.

Guillermo Rauch
@rauchg
X
This is how agents will ship all code.
Nov 14, 2026 · View on X

He names the pieces you can run today as the Vercel CLI commands `vercel flags`, `vercel metrics` and `vercel deploy`.

What is actually being proposed

The interesting part is not the commands, it is where the checking happens. In the usual loop, a human runs the thing locally, pokes at it, and decides whether it looks right before anything reaches users. Rauch's version moves that judgment after the deploy. Code goes out behind a flag, automated gates decide whether the push is allowed, and the signal that the change worked comes from production metrics rather than from someone watching a dev server.

That is a vendor describing its own primitives, and the post is a sketch rather than a result. There is no benchmark here, no before and after, and no claim about how often the gates catch something. Treat it as a proposed shape, not evidence.

Someone already working this way

The shape is not hypothetical. Marc Lou described the same pattern from the other end, as a habit rather than a product pitch. He says he asks the AI to make a feature visible only to him in production so he can assess and iterate, releases it to users if it is good, then has the agent set hourly or daily reminders to check logs and the database and fix issues. His AGENTS.md has rules forcing the agent to test everything.

Marc Lou
@marclou
X
I haven't touched localhost in month.
Nov 14, 2026 · View on X

His summary of the result is blunt.

I haven't touched localhost in month.

Two posts is not a trend, and Lou's account is one person's workflow on his own projects. But the overlap is worth noticing, because Rauch is pitching tooling for exactly the loop Lou says he already runs without it.

For working engineers

Nothing in either post requires Vercel. A feature flag, a CI gate that blocks a push, and something that reads production logs on a schedule are available in whatever stack you already have. The part to copy, if you copy anything, is the decision to make production the place where the change gets verified, with the flag doing the work that a local run used to do. The part to be careful about is that this only works if the gates are real. Rauch lists proof checking and security audits as the gate content, and neither post says what happens when a gate passes something it should not.

Have the good opinion, not the loud one

One email a day. Under three minutes. What working engineers found out about coding with AI.

Get the next one by email

Coding with AI, read daily so you do not have to. The experiments, the receipts and the arguments from engineers shipping real software. Not a changelog.