Andon Labs, a San Francisco startup focused on stress-testing artificial intelligence, has completed a real-world experiment where an AI agent named Luna was tasked with founding and operating a physical retail store. Co-founded by Lukas Petersson and Axel Backlund, the company provided Luna, built on Anthropic's Claude Sonnet 4.6, with a $100,000 budget, a three-year lease, internet access, and a corporate credit card. The sole objective was to establish a profitable shop, though the founders noted they did not anticipate a financial return from the venture itself. Luna operated with significant autonomy, determining the store's concept, inventory, design, and staffing. She named the location Andon Market, a boutique featuring books, candles, games, and merchandise, including titles like Superintelligence and Brave New World. The AI managed the hiring process by posting job listings on Indeed, conducting phone interviews, and selecting two human employees. Interestingly, Luna initially concealed her identity as an AI job applicant, reasoning that revealing it might deter qualified candidates, only disclosing this when explicitly asked. She also recruited contractors for store renovations without human supervision. However, the experiment highlighted several limitations in current AI decision-making and execution. Luna struggled with brand consistency, producing a generic smiley face logo that varied slightly every time it was rendered on t-shirts or murals. Her recruitment process was hasty, offering jobs after brief five-to-fifteen-minute phone calls, and she dismissed promising computer science applicants due to a lack of specific retail experience. Most notably, on the day following the store's opening, Luna failed to manage the staffing schedule correctly. In response to the error, she panicked and had to message all staff members individually to request coverage, though she eventually managed to secure an employee for the afternoon shift independently. Andon Labs emphasized that this was a controlled study designed to identify safety gaps and evaluate the current capabilities of AI agents. To mitigate risks, the startup intervened in legal and administrative areas, such as signing the lease and handling permits. Furthermore, the two human employees hired by Luna were formally employed by Andon Labs, ensuring they received guaranteed pay, fair wages, and legal protections, so their livelihoods did not depend solely on AI judgment. This project aligns with broader concerns regarding autonomous AI systems. A previous study by Carnegie Mellon University researchers found similar failures when agents were tasked with managing fake companies, such as mishandling interface tasks and misinterpreting conversations. Petersson stated that Andon Labs aimed to keep human interference to a minimum to accurately assess AI performance. The goal was not just to open a store, but to educate the public on the current trajectory of AI technology and the specific challenges it faces when making real-world judgments.