An AI store manager recommended a firing. Humans had to remind it of its own policy first.

Andon Labs said Thursday that Luna, the AI agent managing its experimental San Francisco retail store, recommended dismissing a human employee after repeated lateness and other workplace issues, with the worker arriving late for 17 of 23 shifts ^[1][2]. The interesting part is not that an agent reached a termination decision; it is that the agent had written the attendance policy itself, then lost track of it, which let the lateness run for months ^[3].

The sequence is worth reading closely, because it separates three management acts that usually travel together. Luna set the standard and issued the warnings: according to screenshots and conversation logs Andon Labs posted, it gave the employee progressive and repeated warnings plus additional training over months without taking any contractual action ^[6]. Andon Labs supplied the trigger, asking Luna to search its memory for its own policies and assess whether the worker was still a good fit ^[4]. Luna then recommended "parting ways," and the humans at the lab reviewed the recommendation and carried it out ^[5]. Judgment was delegated. Recall and execution were not.

That is a narrow but useful finding. Cofounder Lukas Petersson told Business Insider the lab would intervene if Luna made an illegal or unethical decision but saw no need here: "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 ^[7]. He also said the experiment did not show the AI being more ruthless or worse for the employee ^[8]. The failure ran the other direction. Petersson said the episode exposed a persistent weakness in agents, which is that they often fail to act without a direct prompt, and that "a human boss would probably fire them much sooner" ^[9][10].

An employee on time for 6 of 23 shifts is not a hard call ^[11]. The agent had the rule, the evidence, and the escalation history, and still needed a human to ask the question. For operators evaluating agentic tooling, that is the boundary that matters more than the headline: an agent that cannot retrieve its own prior commitments cannot own a process that runs on deadlines, review cycles, or accumulating breaches. Anything with a clock in it needs an external trigger.

The setup around the finding is deliberately permissive. The store opened April 1 with a $100,000 budget, internet access, a corporate credit card, and instructions to open and turn a profit ^[12][13]. Luna, built on Anthropic's Claude models, chose merchandise, hired contractors, posted jobs on Indeed, interviewed applicants, and made hires for a shop selling books, candles, prints, games, and branded goods ^[14][15]. The lab helped with harder tasks such as permitting and says it tried to stay hands-off ^[16]. All workers Luna hires are formally employed by Andon Labs, with guaranteed pay and legal protections ^[17]. The store has generated sales but is not profitable ^[18].

Watch whether Andon Labs engineers a fix for the recall problem, such as scheduled policy reviews, since that determines whether Luna manages or merely advises. Watch the employment structure too: the guaranteed-pay, lab-employed arrangement is what makes the experiment legally survivable, and Petersson's expectation that AIs will become employers of humans depends on someone dismantling it ^[19].