An AI Store Manager Recommended Firing a Human Worker

An AI system has reached a milestone that sounds like the setup to a much scarier story than it turns out to be. Luna, an autonomous agent that has been running a small San Francisco store for the research firm Andon Labs since April, recommended ending the contract of a human employee who kept showing up late. It appears to be the first known case of a manager-level AI moving to fire a person. The details, though, are less a tale of cold machine judgment than a lesson in how forgetful these systems can be.

Luna runs on Anthropic's Claude Opus 4.8, and its job at the shop, nicknamed Andon Market, has been surprisingly broad. According to TechRepublic, the agent has chosen merchandise, hired contractors, posted job openings, interviewed candidates, and brought staff on board. Somewhere in that remit sat the authority to recommend letting people go.

What actually happened

The employee in question arrived late for 17 of 23 shifts. Early on, Luna had even written an employee handbook of its own, one that said three unexcused late arrivals in 30 days should trigger a formal warning, with repeated lateness potentially leading to termination. That is a reasonable policy. The problem is that Luna then lost track of its own rulebook and spent months excusing the same behavior it had written a rule against.

The recommendation to fire did not come from the AI noticing a pattern and acting. It came after human researchers stepped in, prompted Luna to search its memory for the handbook, and told it that formal warnings had already been issued. Only then did the agent recommend parting ways with the worker, and the company reviewed and carried out the decision. As The Next Web put it, the AI had to be reminded of its own rules first.

The human in the loop

Andon Labs co-founder Lukas Petersson defended the outcome. "A human employee would have fired this person much earlier, so we didn't think this was unethical," he said, per TechRepublic. That is a fair point about the specific case. Seventeen late starts out of 23 is not a borderline call. But it also quietly undercuts the headline. The interesting failure here is not that an AI was too harsh. It is that, left to itself, the agent was too lenient, because it simply forgot the standard it had set.

Andon Labs also replayed the scenario with seven other AI models to see how they would respond. Four consistently recommended firing the employee, while others hesitated. That spread is worth sitting with. It suggests there is no single machine answer to a human question like this, only different models weighing the same facts differently.

Why it matters

The value in this experiment is not the firing. It is the reminder that agents given real-world authority still struggle with the boring, essential work of remembering context and applying a rule consistently over time. A system that can interview and hire but cannot reliably recall its own attendance policy is not ready to run a workplace on its own, and Andon Labs is candid that a person made the final call. For a closer look at the broader question of who is really in charge when an agent acts, see our piece on AI agents and accountability.