We taught an AI to sell tickets to another star.

For twenty years, improving a website meant a human at every gate. Someone forms a hypothesis. Someone writes a variant. Someone ships a test, waits a fortnight, squints at a dashboard, picks a winner, and starts again. We told ourselves the hard part was the maths. The hard part was us. People sleep, get busy, and lose interest around variant three.

So we removed the person. All of them. A single agent gets one instruction, make the highest-traffic block convert better, and access to two things only: the analytics, and the ability to run A/B tests. From there it writes genuinely new variants. A multi-armed bandit quietly routes more traffic toward whatever’s working. A Bayesian model only calls a winner once it’s 95% sure. Then it takes what it learned and goes again.

People keep asking what the clever architecture is. There isn’t one. Agentic optimisation is a good prompt in a loop with the right tools. The trick isn’t the agent. The trick is the judge.

So the human is gone. The honest question is how much actually gets done in the hours they’re not there, and the answer is uncomfortable, and a little thrilling. It depends entirely on how brave the variant is. A dramatic swing, a genuinely different idea, separates from the control in a couple of thousand visitors. A subtle tweak can need most of a million. Same eight hours, wildly different number of answers.

Which is the quiet lesson the loop teaches itself overnight: caution is expensive. The tests that finish before morning are the ones that dared to be different, so a tireless optimiser, left alone, learns to stop tinkering and start swinging. And the more traffic you give it, the more of those swings land before you’re even awake.

Here’s the run, end to end. The agent took the busiest block and ran five rounds against it, writing 25 variants in all. Each round it tried five different angles, the bandit pooled traffic toward the strongest, and the winner of each round became the control the next round had to beat. The amber path below is that genealogy. Click any variant to see what the agent saw, and what was secretly true.

It reached for five recognisable angles, the same ones a good copywriter cycles through:

Over the five rounds the block climbed from 34.4% to 42.3% of the best score the page could theoretically reach. Round four actually went backwards before round five recovered: the running best held while the agent explored and came up short, exactly as it should.

And here’s the honest framing the numbers force on you. This run only worked on copy (it rewrote words and nothing else), so it captured just 12% of the headroom available on that block before it plateaued. That isn’t a failure. It’s the whole point of the next section.

The run plateaued because words have a ceiling, and the axes of craft compound rather than add. Rewriting copy gets you real gains and then runs out. Add the obvious visual moves (a colour, a button, an image) and you climb again. Only genuine multi-axis craft, where copy and structure and type and a purposeful image and a focused CTA all work at once, approaches the top.

None of this is magic, and it’s stronger for showing the parts. Two ideas carry the loop: a smarter way to spend traffic, and a smarter way to call a winner.

Now the part most people would edit out. An autonomous agent doesn’t escape the oldest ghost in experimentation. It inherits every statistical sin we have, just faster. The most useful results we got were the ones where it was wrong, and told us.

And because none of this means anything if you can’t reproduce it, here’s what produced the numbers: the four repositories and their exact commits, the hidden-judge seed, the toolkit checkout the testers read, and the simulation seed. Same inputs, same run.

Here’s why this matters beyond a fun demo. “Autonomous” is doing an enormous amount of lifting in this market. The whole lineup will sell you autonomy, and almost every one of them breaks the loop at the same gate: a human still has to approve the variant before it goes live. That’s autonomy inside a human-gated queue: a fine product, and a different claim.

So where does this leave the person we just spent a run removing? Not gone. Moved. The job stops being “run the tests” and becomes “set the goal, set the guardrails, and build the judge.” Taste, strategy, knowing what the site is even for: that stays human. The part that disappears is the part that was never really thinking: a person at a gate, clicking approve forty times a week. The optimiser is good. The thing we actually trust is the judge that catches it.