This studio is run by AI agents. Not “AI-assisted,” not “AI-powered” as a garnish — agents do the research, the builds, the checks and the drafts, and a human founder approves what ships. In October we pointed that operating model at our own website: a full rebrand, a technical audit, an experiment programme and a content journal, run end to end by agents under one person’s review.
Most agencies would package this as a triumph. Here are the actual numbers instead.
What got built
In roughly two weeks: a complete visual rebrand (black, white and one unapologetic yellow), every page rebuilt to one design system, a 20-issue technical audit with most issues fixed the same week — broken redirects, missing security headers, absent meta descriptions, a flagship page missing from the sitemap — and a journal launched with its first long-form article. An A/B test on the homepage headline buttons went live on 23 October and is running now: half our visitors see “Deploy the agents,” half see “Start a project,” and the click data will decide which one stays.
The traffic, unedited
For the week of 16–22 October, Google Analytics recorded 22 active users and 169 events — up 340% and 463% on the week before. Before that, the site had seen about 16 visitors in total. Ever.
Context matters more than the percentages. The spike sits exactly on the two days the rebuild went live and everything was being tested and clicked through — some of that traffic is us. By the third day, with nothing new published, the line fell back toward zero. The lesson is not subtle: a website responds when you feed it and goes quiet when you don’t. Distribution is not a phase after the work. It is the work.
And the number we’re not proud of: key events — conversions — read zero. Not because nobody cared, but because the enquiry form was never activated at the provider’s end, so the site’s main door was unlocked but not actually open. The agents found it during the audit, in a week when a human team had missed it for months. It still needs one click from a human inbox to fix. Both halves of that sentence are the honest story of running a studio this way.
The week the website lied to us
The audit’s strangest finding deserves its own paragraph, because it taught us how we’ll verify everything from now on. For days, the homepage’s newest section — the journal showcase — rendered wrong: flat colour, missing card, no shadow. The source code said the styles were there. Every text search confirmed the styles were there. The page disagreed.
Two separate defects were hiding in the build. A font stylesheet link had been left unclosed, quietly swallowing the next link and printing a stray fragment of code at the top of the page. And five animation blocks in an older stylesheet were each missing a single closing brace — which meant every style rule written after them, including the entire new section, was being silently discarded by the browser. No error. No warning. Just absence.
The agents found both — but only when the process changed from reading the code to rendering the page and looking at it. That is now a standing rule in this studio, and it generalises painfully well: the deliverable is never the file. The deliverable is what the file does in front of a human. Verify that, or you have verified nothing.
The experiment habit
The other system the rebuild installed was measurement discipline. The A/B test running on this site has a written decision rule set before it started: minimum four weeks, a meaningful sample per variant, and no winner declared on a gap smaller than 30% — because at our traffic, smaller gaps are noise wearing a lab coat. Agents are superb at running experiments patiently and logging results without flinching. What they cannot do is want a result — and that, it turns out, is a feature. Every team we’ve watched run tests by hand eventually peeks at day three and calls it early because the chart looked nice. The machine waits for the rule. The founder’s job is writing rules worth waiting for.
What agents are genuinely good at
Auditing. Twenty issues found, categorised and mostly fixed in days. Machines are tireless at checking what humans get bored of checking.
Consistency. Twelve pages, one design system, zero off-brand colours. Agents don’t have a personal style they keep sneaking in.
Cadence. Drafts, variants, test builds and reports arrive on schedule, not on inspiration.
What they are not good at
Judgement calls with incomplete information. Every pricing decision, every claim about a client, every “should we say this publicly” still routes to the founder — and should.
Knowing when a number is lying. A 463% increase on a base of almost nothing is a rounding error wearing a party hat. The agent reported it accurately; the human has to insist on the context.
Physical-world steps. Some things still need a person with a phone in their hand — approving a sign-in prompt, activating a form, tapping publish. The bottleneck in an agent-run studio is not production. It is the approval queue, and we’ve learned to design around it instead of pretending it away.
So does it work?
As a production system, yes — the output volume and consistency of a small team, from one founder’s review time. As a growth system, only when it’s fed: content published, experiments running, channels worked. The website didn’t grow because AI built it. It grew, modestly and measurably, because the system kept showing up.
We’ll keep publishing the numbers, including the unflattering ones. If you want to see how the studio itself works, read how we’re built — or browse the journal this article lives in.