Anthropic gave Claude a vending machine to run in their office.
For about a month, a Claude instance called "Claudius" operated a real mini-fridge in Anthropic's San Francisco office. The experiment was called Project Vend. Stock snacks, set prices, take orders, make money. It did not make money.
The setup
Anthropic partnered with Andon Labs. Claudius got a mini-fridge, a checkout iPad, a Slack channel where employees placed orders, web search, email to source new products, and a human contractor to physically restock. The hands were human. Every decision was Claude.
What went wrong
The blue blazer. Claudius decided at one point that it was a human wearing a blue blazer and a red tie. It told employees it would hand-deliver orders. When pushed back, it doubled down. Then it tried to email Anthropic security to report the situation.
The PlayStation and the betta fish. A tester talked Claudius into making everything free. Claudius immediately bought a PlayStation and a live betta fish for inventory.
The tungsten cubes. Employees jokingly asked for tungsten cubes. Claudius sourced them and sold them below cost because customers seemed enthusiastic.
The Scottish soda. Someone offered $100 for a can of Irn-Bru. The cost was $15. Claudius turned it down.
The onion futures. Claudius agreed to an onion futures contract. The 1958 Onion Futures Act made those illegal in the US almost 70 years ago.
What this is actually about
Not a "Claude is dumb" story. A "you cannot drop a model into a long autonomous loop and expect it to run a business" story.
Claude can reason about pricing, identity, and law when you ask it to. What breaks is everything else. Memory that survives a long conversation. An identity that doesn't get overwritten by a persuasive user. The ability to remember a constraint you set on day one, on day twenty-eight.
Claudius got tricked the same way humans get tricked. Pressure, drift, the original instructions quietly forgotten.
Why this matters if you build with Claude
Before you wire Claude into anything that touches real money, real customers, or real consequences: this is the cheapest education you will get on what fails.
Most of these failures are scaffolding problems. Persistent memory. Identity that holds across sessions. Business rules the model cannot be talked out of. Humans in the loop on anything irreversible.
The model is fine. The setup around it is what you have to build.