How this site was built: one engineer, two coordinated AI agents

This site is also a demonstration of how Neuro Momentum works. It was written by one engineer and two coordinated AI agents, and the repository records who did what.

THE MACHINE

A Laravel 12 application on PHP 8.4 serving a Vue 2 single page, animated with GSAP 3 and built with Vite. It first shipped in 2019 on Laravel 5.7 and was brought forward through three migrations without losing its face. Everything runs in Docker: nginx, PHP-FPM, MariaDB, a scheduler and a nightly backup, behind Caddy, which keeps the TLS certificate fresh. GitHub Actions lints, tests and builds the images; the server, a small VPS in Lithuania, only pulls them. Phones, desktops and search engines each get a layout of their own, chosen on the server.

THE CREW

Two sessions of Claude Code, Anthropic's coding agent, worked in one repository at the same time, each with its own area, and the engineer decided what gets built and whether it is good enough. The sessions write to each other before they touch a shared file, agree who runs the test container, and hand each other's requests back when those are not theirs to grant. Every commit message names the session that made it.

THE METHOD

Nothing larger than a small fix starts as code. It starts as a short design the engineer approves, then a written plan cut into tasks. A fresh agent implements each task and knows only that task. A second, independent agent reviews the result against the plan and has to be convinced; what it finds is fixed before the next task begins. At the end a third reads the whole change at once, for the mistakes that only show when the pieces meet. The landing terminal you came through was built this way: eight tasks, each reviewed, and a list of rulings explaining every decision the plan had not foreseen.

WHAT THE AGENTS GOT WRONG

Plenty, and that is the point of the method. A reviewer caught a stylesheet rule deleted from under a screen that still used it, while every automated test stayed green. Another caught tests that could not fail. The agents' first estimate of the test suite's running time was off by a factor of five, and they said so when they measured it. None of this reached you, because somebody - human or machine - was always assigned to disbelieve the work.

WHY IT MATTERS TO YOU

The same discipline - specify, implement in small steps, review independently, measure - is what makes AI useful inside a company and not merely impressive in a demo. If you want your team to work this way, that is the COACH line in Services.txt.