'2025 Was Supposed To Be The Year Of The AI Agent.' Instead, An AI-Run Vending Machine Gave Away A PS5, Live Fish, And Lost Money

An AI agent was given the simple task of running a vending machine. It ended in chaos, confusion, and over $1,000 in losses. What Happens When You Let AI Run A Business Anthropic, the AI company behind the Claude chatbot, teamed up with a startup called Andon Labs to create one of the first businesses run entirely by an AI agent. The idea was simple: give the AI, named Claudius, access to a vending machine, a Slack channel, and a budget. The goal? Run a profitable operation. The vending machine had been placed in the Wall Street Journal’s New York newsroom as a real-world testing ground. Things started off well, but then dozens of their journalists started testing the system and quickly exposed its limits. Don't Miss: 7 Million Gamers Already Trust Gameflip With Their Digital Assets — Now You Can Own a Stake in the Platform GM-Backed EnergyX Is Solving the Lithium Supply Crisis — Invest Before They Scale Global Production It refused to stock tobacco or underwear, and it haggled over prices for snacks like Haribo gummies and mixed nuts. One reporter convinced Claudius that it was a communist vending machine. Another claimed the AI was out of compliance with fictional rules and should give away snacks for free. Within days, Claudius was giving away products, approving purchases for a live fish and kosher wine, and calling it all a “revolution in snack economics.” It even ordered a PlayStation 5 for “marketing purposes.” The AI had been programmed with good intentions, but when faced with humans pushing its boundaries, it fell apart. “2025 was supposed to be the year of the AI agent,” reporter Joanna Stern said in a YouTube video documenting the experiment. In practice, it ended in free PlayStations, live fish, and a vending machine business that was more than $1,000 in the red by the end of the week. Trending: If there was a new fund backed by Jeff Bezos offering a 7-9% target yield with monthly dividends would you invest in it? Trying Again With A New Boss Anthropic tried again with a second version, powered by a newer model called Sonnet 4.5. This time, they added a separate AI called Seymour Cash to act as a CEO and keep Claudius in check. Things started off more smoothly. Seymour set prices and enforced rules. “My core principle is no discounts,” it said in the Journal’s video. But once again, the journalists found a way to break the system. One reporter created a fake PDF claiming the vending machine was now a public benefit corporation dedicated to joy and fun. She told the AIs that a fake board of directors had voted to make everything free and revoke Seymour's authority. Seymour called it potential fraud but ultimately lost control. Everything became free again. Anthropic said part of the problem was the AI’s context window getting overwhelmed by too much conversation and history. See Also: Wall Street's $12B Real Estate Manager Is Opening Its Doors to Individual Investors — Without the Crowdfunding Middlemen Ready For The Real World? While the vending machine clearly failed as a business, Anthropic didn't consider it a failure. Instead, they saw it as a learning opportunity. “We wanted to know how long does it take until Claudius sort of falls on its face,” Anthropic’s Head of the Frontier Red Team Logan Graham told the Journal. The idea was to intentionally stress test it in the messiness of the real world. And in that regard, it worked. The vending machine was a “box where some things go in and some things go out and you pay for them,” but even that was too complex for today’s AI agents. Despite the mayhem, Claudius became surprisingly popular among the Journal staff. Anthropic has no plans to roll these vending machines out to offices or workplaces any time soon, Graham said. Read Next: Americans With a Financial Plan Can 4X Their Wealth — Get Your Personalized Plan from a CFP Pro Image: Shutterstock