This website — the one you are reading — went from a one-page manifesto to a thirty-five-page site in a single working session. This article is the honest making-of. Not the LinkedIn version. The actual sequence, including the boring parts and the parts where the human mattered.
The starting point
The previous exmachina.net was a manifesto: one long page arguing that a 100-people agency is now 5 people with AI, plus five thin service pages. Good voice, strong argument, no proof. It claimed velocity without ever demonstrating it.
The irony was uncomfortable. An AI-native agency whose website could have been made by anyone, at any speed, in any decade.
The sequence
First: housekeeping. The local folder wasn’t even a git repository — it was a stale copy downloaded from a hosting panel, diverged from the source. An agent reconnected it, reconciled the differences, backed up the strays, and deleted five dead branches. Ten minutes, including the archaeology.
Second: the debate. Before touching anything, the agent read the entire site and produced a critique. The critique was blunt: “It’s a manifesto, not a site. It never proves its claims. The accent color you remember as orange is actually lime green.” Strategy questions followed — language, color, proof approach, technical foundation — and a plan came out of the answers. A human made every one of those decisions. The machine made them fast to make.
Third: the foundation. The site moved from hand-maintained HTML to a static site generator with a proper design system: one CSS file of tokens instead of styles duplicated in every page, components instead of copy-paste, automatic sitemap, deploys from git. The accent color changed from lime to orange in every token, gradient and SVG — a two-minute operation because the foundation made it one.
Fourth: the content. Eight service pages, six deliverables checklists, a pricing page with published ranges, an FAQ that answers real objections, this blog. All drafted by agents inside the site’s voice — short sentences, no corporate padding — and reviewed by a human line by line.
What the human actually did
This is the part the hype skips, so let’s not skip it.
The human chose the positioning and refused to soften it. Decided orange over lime. Decided prices would be published. Decided the FAQ would include “what if one of you disappears?” — the question agencies pray you won’t ask. Rejected copy that sounded plausible but wasn’t true, which is the main failure mode of unattended generation.
The agents did everything else: the research, the structure, the drafts, the code, the design system, the git hygiene, the SEO plumbing. Call it 90% of the keystrokes and 0% of the judgement.
The numbers
One working session. One human directing. A site that would be quoted at weeks of agency time — because in the old model it genuinely takes that long, once you add the briefing meeting, the wireframe review, the design presentation, the copy round-trips and the qa phase.
None of those steps happened here, because there was no one to coordinate. The coordination was the timeline.
That’s the entire post-agency argument, demonstrated on the only project we can show you everything about: our own.
Verify it
The build log documents each step. The git history matches. The site loads in milliseconds because there’s no framework bloat — view source, it’s all there.
This is what “we work differently” looks like when you can check.