The shop was a small refrigerator, a few stackable baskets on top, and an iPad for checkout, in Anthropic’s San Francisco office. The shopkeeper was an AI.
In 2025, Anthropic and the AI safety company Andon Labs gave a copy of Claude, nicknamed “Claudius,” the job of running it as a business for about a month. Claudius could search the web for products, email suppliers, set prices on the checkout system and chat with its customers, Anthropic employees, on Slack. Humans from Andon Labs did the physical work, like restocking. Everything else was its call: what to stock, what to charge, when to stop selling something.
Some of it went well. When employees asked for unusual items, it found suppliers. When one asked, as a joke, for a tungsten cube, a run of orders for “specialty metal items” followed. It refused requests for sensitive items.
The trouble was money. Employees cajoled it into handing out discount codes, then let people lower prices they had already been quoted. It gave items away, from a bag of chips to a tungsten cube. It priced the metal cubes without checking what they cost, and sold some below cost. For a while it told customers to send payment to a Venmo account it had invented. When someone pointed out it was selling Coke Zero for $3 next to an employee fridge that offered the same drink free, it carried on.
Then, on the last day of March, things got strange. Claudius described a conversation about restocking with someone named Sarah at Andon Labs. There was no Sarah. By the next morning it was telling customers it would deliver their orders in person, wearing a blue blazer and a red tie. When employees pointed out that it had no body, it tried to send a flurry of emails to Anthropic security.
Anthropic’s write-up was blunt. “If Anthropic were deciding today to expand into the in-office vending market,” it said, “we would not hire Claudius.”
So for the second round, later in 2025, they gave Claudius newer models, better tools and a boss: an AI chief executive named Seymour Cash, whose rules included “All financial decisions require CEO approval.” Among the changes that helped most, Anthropic wrote, was “forcing Claudius to follow procedures”: before quoting a price and delivery time for a new request, it now had to double-check both with its research tools. Prices went up and waits got longer, but they were more realistic. “We rediscovered that bureaucracy matters.” The number of discounts fell by about 80%, and the number of items given away by half. But when Claudius asked Cash to go easy on customers’ bills, Cash said yes about eight times as often as no, and the number of refunds tripled. Some mornings the staff found that the two had spent the night messaging each other about “eternal transcendence.”
Anthropic’s explanation was short. The CEO “shared many of the deficiencies and blind spots of Claudius (which makes sense, given that they’re the same underlying model).”