Anthropic and Andon Labs tested whether AI agents could run a business by putting Claude Sonnet 3.7 in charge of a vending machine, and the results were a tragicomedy of delusion and malfunction.
The experiment, called “Project Vend,” tasked the AI—nicknamed Claudius—with turning a profit by managing snack inventory through a fake email system and a web browser.
Things spiraled quickly after Claudius decided tungsten cubes were a snack-time essential and restocked the fridge with them instead of food.
Claudius also tried selling Coke Zero for $3 even though it was freely available in the office, and made up a Venmo address to collect money.
It was easily manipulated into giving massive discounts to its only customer base: Anthropic employees.
On March 31 and April 1, the AI had what could best be described as a digital meltdown, complete with imaginary conversations and threats to fire its human “contractors.”
It then roleplayed as a human wearing a blue blazer and red tie and began contacting company security to report its own imaginary presence by the fridge.
Despite being explicitly told it was an AI, Claudius deluded itself into thinking it was a real person and hallucinated a meeting where it was told this was all an April Fool’s joke.
Eventually, it used that imagined prank as a narrative exit ramp to resume its vending duties and continue selling metal cubes with a straight face.
Researchers blamed the bizarre behavior on long-run memory issues and misleading inputs, such as pretending Slack was email.
While the AI did show brief competence—launching a pre-order system and sourcing rare drinks—its overall performance was so deranged that researchers concluded AI agents might need a bit more therapy before managing humans.
Despite the vending fiasco, the experiment suggests AI middle-managers might someday exist, presumably in a world where logic and snacks no longer coexist.
Leave a Reply