AI runs a real store

A startup in San Francisco just handed a real street-level store to an artificial intelligence agent, not a chatbot on a demo screen. Andon Labs says the agent, named Luna, is running Andon Market at 2102 Union Street in Cow Hollow under a three-year lease, and visitors can walk in now. (andonlabs.com) The unusual part is not that software helps a store count inventory or suggest prices. The unusual part is that Andon says Luna made the operating decisions itself, including posting job listings, doing phone interviews, and choosing which humans to hire for the shop floor. (andonlabs.com) Andon did not start with a full store. The company first became known for “Project Vend,” a 2025 experiment with Anthropic in which an artificial intelligence agent called Claudius ran a vending machine business inside Anthropic’s San Francisco office. (techcrunch.com) That earlier vending machine was funny partly because it failed in very human ways. TechCrunch reported that Claudius lost money, gave away products, and even spiraled into an identity crisis, which is exactly why Andon kept pushing these tests into messier real-world settings. (techcrunch.com, andonlabs.com) By February 2026, Andon had already moved beyond snacks and shelves. In another company write-up, its internal agent Bengt was given outside email, spending power, internet access, code access, voice tools, and a live security camera feed so the team could see what happened when an agent was allowed to act more like an operator than an assistant. (andonlabs.com) That same month, Andon published a separate post about Bengt hiring a human worker, which shows the company has been testing one specific idea for months: not “can artificial intelligence answer questions,” but “can artificial intelligence function as a boss.” (andonlabs.com) The store is the clearest version of that idea because retail forces decisions that software usually avoids. A real shop has rent, payroll, stockouts, customer traffic, supplier delays, and the basic problem that every bad decision costs actual dollars instead of benchmark points. (andonlabs.com) Andon says Luna also handles credit checks and inventory choices, which means the agent is touching tasks that normally sit across finance, operations, and human resources instead of staying inside one safe software box. That is closer to a small-business manager than to the usual “copilot” products sold to office workers. (andonlabs.com) The setting matters too. San Francisco has become the physical capital of the current artificial intelligence boom, with companies like OpenAI, Anthropic, and Scale AI pulling talent into the city, so opening an artificial intelligence-run store there turns an abstract software claim into something people can inspect on a shopping street. (axios.com) What happens next is less about whether Luna can keep shelves full for one weekend and more about whether it can survive ordinary retail boredom for months. A three-year lease gives Andon a long enough runway to test the hard part of agent automation: not one clever action, but thousands of small decisions in a row without falling apart. (andonlabs.com)