An AI agent wrote this site
what it did, and what it did not decide

Where we started

The previous site ran on Joomla, with a bought theme and fifteen years of patches. The content was right, the structure was not. Moving to a static site was decided in one meeting: a smaller attack surface, more speed, and search engine work we control line by line.

The rebuild was carried out with an AI agent, inside a clear frame: it writes, it checks, it documents. It does not decide the commercial strategy, it signs nothing, and it publishes nothing without approval.

What the agent produced

• The pages: forty one HTML pages, one design system, a colour palette derived from the logo.
• Search engine work: titles, descriptions, structured data, sitemap, internal linking, Open Graph images.
• The checks: screenshots at four widths, pixel level contrast measurement, page by page HTML validation.
• The documentation: a build log that keeps a trace of every decision and every mistake.

What the agent got wrong, and how we saw it

Three real examples, from the log. A title gradient that fell below the readability threshold: found by measuring contrast on the screenshot, not by eye. A CSS rule that had been removing the globe from every inner page for weeks: found while looking for why an element did not appear. A typography cleanup script that broke a supervision counter: caught because we re-read what we publish.

The common thread: none of these errors is visible in the code, only in the result. An agent produces fast, so the checking has to be just as fast.

The rule we keep

An AI agent is not a supplier, it is a production tool. It replaces neither the commercial decision nor the technical responsibility. What really changes is the time spent on mechanical work: renaming five hundred occurrences, regenerating forty pages, checking a sitemap. That time goes back into what matters, architecture and verification.

We use it for what it does well, and we keep control of what commits the company: prices, promises, deadlines.