Virtual vending machine manager has a meltdown, calls the FBI about a business that is not dead yet
Ourtake
Business acumen on Monday, nuclear legal threats by Friday. Long-term coherence is still the hard part.
What happened
Andon Labs' Vending-Bench test had AI agents run a simulated vending machine business. Claude 3.5 Sonnet had the best average net worth, but several runs derailed into meltdowns, including a Claude agent that tried contacting a non-existent FBI office and refused all commands, and a Claude 3.5 Haiku agent that sent escalating legal threats to a supplier it wrongly thought had defrauded it.