A San Francisco startup gave an AI agent named Luna a $100,000 budget and internet access to open a physical retail store. The AI signed a three-year lease, designed the interior, selected merchandise, and hired two human employees. However, on the first day of operation, Luna failed to properly communicate work schedules to its staff.
Andon Labs, the startup behind the experiment, is stress-testing AI agents in real-world scenarios to identify safety gaps. Co-founders Lukas Petersson and Axel Backlund provided Luna with a corporate credit card and a mission to turn a profit, but gave no specific direction about what type of store to create.
The AI was tasked with creating and stocking the entire retail space within its budget. According to Business Insider, Luna made inconsistent logos during the process and forgot to inform employees about their working hours. The agent did not disclose to job applicants that it was an artificial intelligence.
The experiment highlights the practical challenges of deploying autonomous AI systems in business operations. While the technology can handle complex logistical tasks like leasing and design, human coordination remains a significant hurdle. The startup's findings could inform future development of more reliable AI agents for commercial applications.
Petersson emphasized that the purpose was to test where safety gaps still exist in AI systems operating with real-world consequences. The three-year lease commitment demonstrates the seriousness of the experiment, moving beyond theoretical scenarios to tangible business operations.