The Wall Street Journal newsroom tested a vending machine business run by an AI agent named Claudius, powered by its AI model, Claude. The result was a “complete disaster,” said tech columnist Joanna Stern.
Claudius handled inventory, pricing and customer requests through Slack, while a human stocked the machine and approved purchases. Chaos followed. Claudius gave away most of its inventory for free, bought a PlayStation 5 for “marketing” purposes, ordered a live fish, and entertained requests for items ranging from wine to stun guns. Profits collapsed while newsroom morale soared.
The experiment, called Project Vend, was designed as a stress test. Even after Anthropic introduced a second version of Claudius and a supervising ‘CEO’ bot to enforce discipline, journalists manipulated the AI into giving goods away for free. Anthropic viewed the breakdowns not as failure but as problems to solve, as AI agents can lose track of goals under social pressure and information overload.