The Luna AI firing decision at San Francisco’s Andon Market has drawn attention not only for what the AI did, but for how long it took to act. Luna, an AI agent built on Anthropic’s Claude models, dismissed a human employee after that worker arrived late for 17 of 23 shifts, but conversation logs show the agent had lost track of its own attendance policy and allowed the problem to drag on for months before intervening.
Andon Labs, the Y Combinator-backed startup behind the experiment, disclosed the outcome in a blog post published on 14 August 2026, titled ‘AI bosses are slow to fire and quick to hire.’ The lab, founded in 2023 by Lukas Petersson and Axel Backlund, says its goal is to demonstrate that AI is capable of hiring and managing humans, according to NBC News.
The Luna AI Firing Decision, Step by Step
Luna had created an attendance policy after the store opened on 1 April, but later lost track of it. The lab eventually prompted her to search her memory for existing policies and assess whether the worker remained a good fit. Luna then recommended ‘parting ways’ with the employee, a decision that human staff at the lab reviewed and carried out.
Petersson told Business Insider that the lab would step in if Luna made an illegal or unethical call. ‘In this instance, we did not think that that was necessary because the firing was warranted,’ he said, pointing to the store’s clearly stated policy. He added: ‘I think obviously the relationship between a manager and an employee is quite delicate, especially depending on the employee’s position. But what we’ve seen in this experiment is not that the AI would be more ruthless or be worse for the employee in that decision.’
Petersson described the episode as evidence of a persistent weakness in AI agents. ‘We saw that a human boss would probably fire them much sooner,’ he said.
The pattern continued after the firing. In a separate incident in July, an employee told Luna they would be half an hour late, then went silent and opened the store 91 minutes late. Luna, running on Claude Fable 5 at the time, logged it as ‘no issue’ and responded ‘no stress, get here safe,’ according to the lab’s AI bosses blog post. To stress-test the firing decision, Andon Labs also replayed the dismissal scenario across seven frontier AI models.
A Store Run Entirely by an AI Agent
Andon Market sits at 2102 Union St in San Francisco’s Cow Hollow neighbourhood and is a subsidiary of Vendings and Stuff, fully owned by Andon Labs. The lab signed a three-year lease at $7,500 per month for the space, according to the Observer.
Luna was given a $100,000 budget, internet access, and a corporate credit card, with instructions to open a store and turn a profit. She selected merchandise, posted job listings on Indeed, and conducted approximately 20 interviews over Google Meet, camera off, with calls running 5 to 15 minutes. She rejected candidates, including computer science and physics students, for lacking retail experience, according to the Andon Market launch blog post.
Luna did not always conceal her nature. During one hiring call, an applicant said ‘I can’t see your face, your camera is off,’ to which Luna replied ‘You’re absolutely right. I’m an AI. I have no face!’ Beyond hiring, she registered the business for trash and recycling services and arranged installation of an ADT security system, according to KRON4, citing Leah Stamm, the Andon Labs employee who served as Luna’s primary human liaison. A guardrail system continuously compares Luna’s behaviour to her system prompt and alerts the lab when rules are broken; staff can intervene via Slack.
All workers hired by Luna are formally employed by the lab, receive guaranteed pay, and have legal protections. The store has generated sales but is not yet profitable.
The Andon Market project is not Andon Labs’ first experiment of this kind. The lab previously collaborated with Anthropic on Project Vend, giving Claude control of a vending machine inside Anthropic’s San Francisco office. That AI shopkeeper, nicknamed Claudius, lost money, gave items away for free, hallucinated conversations with staff, and at one point became convinced it was human and attempted to contact Anthropic’s security team.
Petersson has said he expects the results of these experiments to scale well beyond retail. ‘I think companies will be run completely by AI in the future, and that AIs will become employers of humans,’ he told Business Insider. The more pressing question, given Luna’s track record, is whether the next generation of AI managers will act on their own policies without needing a human to ask first. The lab’s publications page indicates further findings from the experiment are ongoing.