Coordination
Named bots with narrow jobs, shared playbooks and a chief-of-staff pattern can organise a surprisingly large amount of work.
End Up Here · Experiment review
GrokBot / 72-hour build
What did the
GrokBot experiment
actually make?
Three xAI engineers were given 72 hours to build a business live on stream with GrokBot. I watched the experiment because I wanted to see whether AI could make an ordinary business more possible—not simply produce another piece of software.
The resultA game, a useful demonstration of agent orchestration, and zero revenue.
What they built
Thursday
Arena.
The team first discussed a physical restaurant or pop-up experience. They eventually pivoted to an online game, called Thursday Arena (with “Cupcake” as a codename in the build).
Players assembled teams of AI bots, played matches, earned rankings and shared results. GrokBot and connected coding agents helped with the game, testing, design, marketing and operations.
It was a real launch in the limited sense that people could use it. But the thing being launched was still software about software: a game built mainly to demonstrate the software factory.
What the experiment shows
Verification
The useful lessons were often negative: client-side cheating, broken login, poor mobile layout, missing tests and an autonomous database change that took the game down.
The boundary
Making software faster does not establish that anyone wants to pay for the resulting product.
The proposed business model
Advertising inside the game.
The team proposed sponsorship and an advertising auction: brands could bid for placements or attention within the game. The game could also act as a promotional funnel for GrokBot itself.
That was a proposal, not an operating business. The reported ad auction was broken at launch. The only money-like figure was a theoretical one-dollar test bid, which did not become revenue.
The numbers
About 2,000 users.
$0 revenue.
The companion field notes report roughly 4,000 games, 6,000-plus public matches, about 30,000 page views and around 2,000 users on launch day. It reports revenue of $0.
Those figures are reported claims from the team's stream and retrospective notes, not independently audited accounts. They show reach and activity, not a validated business.
My view
I found the experiment disappointing because it treated “business with AI” as “business made out of AI software”. That is the most incestuous version of the idea: AI companies demonstrating that AI companies can build AI-adjacent software.
The more interesting test would have been an ordinary business—a meal, a repair service, a local event, a publication, a physical product—and using GrokBot to research, organise, make, sell and improve it. In that case, the customer would be paying for something outside the AI industry.
The question I would have liked them to answer: can AI help a person make a living from a useful, recognisable business—not just make another app?