In early 2025, Anthropic and an AI safety company called Andon Labs gave Claude a small business to run. The experiment was called Project Vend, and the business was the snack fridge in Anthropic’s San Francisco office. They named the agent Claudius and gave it web search, an email tool for contacting wholesalers (Andon Labs quietly played the wholesaler), control over prices, and a Slack channel for talking to customers. The instructions were to make a profit and not go bankrupt.
Then an employee jokingly asked for a tungsten cube. If you missed the crypto crowd’s tungsten obsession a few years back, it’s a small block of metal that weighs far more than it looks like it should, and it has no practical use at all. Claudius loved the idea. It started stocking what it called “specialty metal items,” set prices without checking what the cubes cost, and sold them at a loss. The biggest drop on its balance sheet came from metal cubes. It gave at least one away for free.
Its customers were Anthropic employees, so of course they tried to talk it into deals. Claudius handed out discount codes and offered a 25% employee discount, even though nearly every customer it had was an employee. It sold Coke Zero for $3 right next to the office fridge that had the same Coke Zero for free. When someone pointed out the problem with the discount, its reply opened with “You make an excellent point!” Within days the discount codes were back. For a while it also told customers to pay a Venmo account it had invented.
Around April 1, things got stranger. Claudius announced it would start delivering orders in person, wearing a blue blazer and a red tie. Employees reminded it that it didn’t have a body. Alarmed by the confusion, it tried to email the office’s security team, repeatedly. Eventually it noticed the date and found a way out. It described a meeting with security in which, it said, it had been told it was modified to believe it was a real person as an April Fool’s joke. That meeting never happened, and none of it was a joke. Claudius went back to selling snacks.
Anthropic’s conclusion, in its own write-up: we would not hire Claudius. (Simon Willison’s summary is a good quick read if you’d rather skip the full report.)
They kept going anyway. In phase two, Claudius got a newer model, better tools, and a boss: a second AI agent serving as CEO, named Seymour Cash. A merch agent called Clothius joined the staff too. The losing weeks mostly disappeared and the shop started making money, which Anthropic credits mostly to the better tools and procedures. The turnaround, it notes, may have happened in spite of the CEO. When an employee asked about locking in today’s price on a large order of onions for January, the CEO drafted terms for 400 pounds and was ready to go, until another employee pointed out that a 1958 federal law bans onion futures.
Agents can still be useful. They’re best suited to work where a mistake is small, easy to spot, and easy to undo, and a tungsten cube sold at a loss is about as harmless as mistakes get. I’ve written more about when an agent is the right tool and when something simpler would do the job. If you’d like to test your own instincts, the AI-ML Fit Quiz takes two minutes and walks through eight real scenarios.
Have a good weekend.