Behind the scenes
This site was rebuilt overnight by AI agents
Shortly after one in the morning on 15 October, I recorded a long, rambling voice note for Codex, the app I work with agents in. Roughly: find my old website, find the feedback I gave on a redesign last month, and rebuild the whole thing around what I want to do now, which is helping people and companies put AI to work. Then I left it alone for the night.
This is what happened, taken from the chat records.
How the work was split
One chat ran the project. It used Claude Opus 5.5 and kept the decisions that need judgment: what the site should say, how it should look, and what was good enough. The heavy lifting went to GPT-6 Astra, OpenAI’s model, in separate chats that I can open in the Codex app and read line by line. Each chat got one job, a list of files it owned, and the checks it had to pass before it reported back.
That split mattered more than the choice of models. Two agents editing the same file at the same time make the same mess two people would. So the two chats that built pages each worked in their own copy of the code, and the lead merged their work afterwards.
The night, from the chat records
Times are Berlin time, 15 October 2026.
- Faruk
Records the brief as a voice note and leaves the agents to it.
- Lead · Claude Opus 5.5
Finds the code on GitHub, the live project on Vercel, and last month’s private draft with my feedback in its notes.
- Platform · GPT-6 Astra
Starts moving the 2024 site off a CMS that no longer runs and onto the current version of Next.js.
- Research · GPT-6 Astra
Starts checking what today’s agents can actually do, and how other people sell this kind of work.
- Research
Done, with 44 sources. The opening line on the home page changes because of it.
- Platform
Done: new platform, old addresses redirected, first preview deployed.
- Lead
Hands the new copy and a design brief to two build chats, each working in its own copy of the code.
- Review · GPT-6 Astra
Reads the copy as three imaginary visitors and lists everything that sounds false, vague, or unfair.
- Services build · GPT-6 Astra
Done: the services and contact pages, checked at four screen widths.
- Home build · GPT-6 Astra
Done: the home page, about, writing, the new favicon, and the share image.
- Lead
Merges both builds and looks at every page in a real browser.
- QA · GPT-6 Astra
Starts the last pass: the readiness checker, this log, accessibility, speed, and every page at seven screen widths.
- QA
Done: no serious accessibility issues left, mobile speed scores of 90 and up, and a fresh preview deployed.
- Lead
Edits the last old article into plain language and writes up the decisions left for me.
Where it got stuck
- The platform chat tried to sign in to Vercel to look at the preview in a browser and ran into two-factor authentication. Nobody was awake to approve it. The lead told it to test the production build on the machine instead and leave the login for me.
- The first attempt to screenshot my old draft failed because the lead sent its access token in the wrong header. It looked up the documentation and tried again.
- A background install quietly did nothing. The build chats ran their own.
- The review chat found real problems. The comparison on the home page made ChatGPT look worse than it is, and a few lines implied a track record I don’t have yet. The lead fixed both. It also ignored some suggestions that would have made the text vaguer.
- My old blog posts from 2023 and 2024 lived in a CMS that is offline now. They are not part of the new site, and their old addresses redirect to the new pages.
What was left for me
The agents did the typing, the checking, and a good share of the thinking. They did not decide what goes live. In the morning there was a preview link and a short list of decisions:
- whether the headline should be “Put AI to work,”
- how long engagements should be, and whether to promise fixed prices,
- whether to say that I live in Frankfurt,
- whether to keep an older article that no longer fits,
- and when to put it all live.
What I take from it
None of this needed a lab or a big budget. It needed a clear brief, one owner for each piece, checks that run against the real site instead of trusting a report, and a person who decides. That is the same work I do with clients, pointed at their jobs instead of my website.
The models and tools named here are the ones I used on 15 October 2026. They change quickly.