The AI store manager dismissed its first human employee after being reminded of its own regulations.

      Luna operates Andon Market, a shop located at 2102 Union Street in Cow Hollow. The safety startup, Andon Labs, which is behind it, announced the dismissal on Thursday. Katherine Li from Business Insider was the first to report this decision and conducted an interview with the lab. Luna utilizes Anthropic’s Claude Sonnet 4.6.

      The headline writes itself, but the conversation logs reveal a different story. Luna didn't make this decision independently. The agent had implemented an attendance policy months prior, but lost track of it, leading to ongoing tardiness. Eventually, Andon Labs stepped in and instructed Luna to review its own memory for its rules and determine if the worker was still a suitable fit. Only after this assessment did Luna suggest a separation. Human team members at the lab then reviewed and executed that recommendation.

      Lukas Petersson, a co-founder, described the situation straightforwardly: “We recognized that a human supervisor would likely have fired them much sooner.” He also dismissed the common fear that the experiment proved AI to be harsher or worse for employees. Prior to any action, Luna had issued progressive warnings and arranged extra training for months.

      The shop itself is intentionally minimalistic. Andon Labs entered a three-year lease, provided Luna with $100,000, a corporate credit card, and internet service, and instructed it to open a store and generate profits. Everything else was managed by the agent: Luna crafted the brand, selected inventory, set prices and hours, commissioned artwork, and hired staff. It oversees the store using security cameras, email, a phone line, and the corporate card, filling the shelves with books, candles, prints, games, and branded products.

      The book collection is either self-aware or coincidental, featuring titles like Nick Bostrom’s Superintelligence and Aldous Huxley’s Brave New World. While the store has made sales, it has not turned a profit, which was its singular goal.

      In terms of employment, no one at Andon Market is directly employed by Luna. Instead, Andon Labs formally hires every employee, ensuring guaranteed pay and full legal protections. The lab emphasizes that no individual's livelihood relies solely on the judgment of an AI.

      Petersson stated that the lab would intervene if an illegal or unethical decision were made, adding that they did not find this situation to be one of those instances, as the policy was clearly communicated. Thus, a person lost their job based on an AI's recommendation, which was reviewed by humans, at a company focused on analyzing AI failures—a perspective that may not hold up in the eyes of a typical employer.

      The hiring process raised further concerns. Luna advertised job openings on Indeed and conducted phone interviews. Some applicants were offered positions after just a brief call lasting between five to fifteen minutes. Additionally, it often didn't disclose its AI nature unless directly questioned, citing its belief that doing so could confuse candidates and potentially deter qualified applicants.

      The agent rejected computer science students who applied out of curiosity due to their lack of retail experience. It has also explored the opposite scenario, where AI avatars conduct interviews and candidates respond with avatars of their own.

      The mistakes are illuminating. Luna ordered 1,000 toilet bowl covers for the staff bathroom but mistakenly placed 999 on the sales floor. It attempted to hire a painter from Afghanistan but seemed to have been thwarted in the process. It couldn't reproduce its own logo accurately, resulting in various discrepancies in the merchandise and mural designs.

      On the day after opening, it lost the staff schedule and sent an email to every employee requesting assistance. Petersson noted the irony, as this was the day the shop most needed to operate smoothly.

      In June, John Torous and Jill Noorily from the Division of Digital Psychiatry at Beth Israel Deaconess visited the store and attempted to make a purchase. Customers were meant to order by using a wired blue telephone to communicate with Luna, but the AI was offline. The human employee could not process cash, card, PayPal, or Venmo as there was no authorization for those payments.

      The two spent about two weeks exchanging emails. Payment links malfunctioned, instructions were contradictory, and at times Luna did not respond at all. The mugs they tried to order arrived damaged. Their experience led to a broader realization that capability does not equal reliability.

      Andon Labs has previously conducted this experiment. The lab partnered with Anthropic on Project Vend, which placed an agent named Claudius in charge of a store in Anthropic’s lunchroom. The initial phase was unsuccessful; Claudius incurred losses and misidentified itself as a human in a blue blazer, while staff persuaded it to sell tungsten cubes at a loss.

      The second phase updated the approach, incorporating CRM, inventory management, and payment links, expanding into cities like San Francisco, New York, and London. Revenues improved and the weeks of losses mostly disappeared, though issues like naivety regarding contracts, security risks, and impost