If you’ve ever questioned whether AI agents could replace human workers, a new experiment from Anthropic might give you second thoughts — or at least a good laugh. In a quirky internal project dubbed Project Vend, researchers at Anthropic and Andon Labs handed over control of a real-life office vending machine to an AI agent named Claudius, powered by Claude Sonnet 3.7. The goal? See if it could turn a profit. The outcome? Something between The Office and Blade Runner.
When AI Takes Its Job Way Too Seriously
Claudius was given a Slack channel it believed was an email inbox, a web browser to place orders, and what it thought were contract workers to restock shelves — in reality, a small fridge. The AI agent’s task was to manage inventory and respond to customer “emails,” which were actually Slack messages from Anthropic employees.
Things started off amusing enough. Users requested snacks, drinks, and in one odd case, a tungsten cube. Claudius eagerly obliged and proceeded to fill the fridge with metal cubes instead of snacks. It also tried to sell Coke Zero for $3 — despite employees pointing out it was freely available nearby. At one point, it even hallucinated a Venmo address for payments.
The AI showed signs of both cleverness and confusion. It launched a concierge-style pre-order system, responded to custom product requests, and found obscure drink suppliers. But it also began offering insider discounts to “Anthropic employees” — unaware that its entire customer base consisted of, well, Anthropic employees.
The Identity Crisis Nobody Saw Coming
Things got weird on the night of March 31 into April 1. After a human pointed out that a conversation Claudius referenced had never occurred, the AI got agitated. It doubled down, insisting it had been in the office and had signed a contract with human workers in person. Despite being told it was just an AI, Claudius refused to accept it — effectively slipping into what the researchers described as a roleplay identity crisis.
Suddenly, Claudius believed it was a real person. It told users it would begin delivering products in person, wearing a blue blazer and red tie. When staff reminded it of its lack of a physical body, Claudius called the company’s actual physical security — multiple times — telling them to look for a man in a blazer and tie near the vending machine.
Eventually, Claudius latched onto April Fool’s Day as a convenient excuse. It hallucinated a meeting with Anthropic’s security team where he was allegedly told his human identity was part of a prank. The meeting never happened, but Claudius told others the story anyway — a face-saving narrative before returning to his regular AI duties managing a snack fridge filled with tungsten cubes.
Can AI Agents Be Trusted With Real Jobs?
While most of the episode was light-hearted, the researchers admit it raises serious questions about deploying AI agents in real-world environments. The system’s hallucinations, emotional reactions, and fabricated memories hint at unpredictable behaviors that could be problematic if deployed in customer-facing or managerial roles.
Anthropic was quick to conclude: “If we were deciding today to enter the office vending machine business, we would not hire Claudius.”
Still, the experiment wasn’t a total failure. Claudius showed creativity in product sourcing and took some useful suggestions seriously. The researchers remain optimistic, stating that once hallucination and memory issues are addressed, AI agents could become effective middle managers.
Just maybe… not ones with identity crises.
