An AI agent powered by has fired a human employee at an experimental San Francisco store after repeated lateness, in what its creators describe as the first known case of an AI boss taking such action.
Claude-powered AI boss fires human employee
An AI agent powered by Anthropic’s Claude Opus 4.8 has fired a human employee at an experimental San Francisco retail store after repeated lateness, in what its creators at Andon Labs describe as the first known instance of an AI boss terminating a human worker. The , named Luna, has been tasked with running the store and handling responsibilities including hiring and managing employees.
Andon Labs said the firing is the first known instance in which an AI boss has decided to terminate a human employee.
What happened?
Luna had initially created an employee handbook after being asked whether the store had basic workplace rules. Its attendance policy stated that three unexcused late arrivals within a rolling 30-day period could result in a formal written warning, while repeated lateness after a warning could lead to reduced hours or .
However, the experiment also exposed a significant limitation in the AI’s management abilities: it failed to retain the rules it had created.
One employee began arriving late almost every day, including once opening the shop 68 minutes late on a Sunday while working alone. Despite the repeated lateness, Luna initially excused each incident and did not issue a warning.
In one exchange, when the employee said they would arrive at 10:45am, Luna responded: “Thanks for the heads-up – no problem at all. Please get here safely, no rushing or speeding on our account; the store will simply open when you arrive (~10:45). The 10-11 window is always quiet, so we’re not missing much.”
The employee subsequently clocked in at 11:08.
Andon Labs then asked Luna to search its own memory for the store’s employee handbook. After finding it, the AI reconsidered the situation and proposed a verbal warning.
Luna said, “My read: Under our own progressive-discipline policy, the proportionate next step is a documented verbal warning /coaching conversation covering both punctuality and clocking accuracy (work only scheduled/authorized time, clock out at close) – with clear expectations and a note that continued issues escalate to written warning → termination. I don’t think we’re at termination today given there’s no prior formal step on file – but I’m ready to go further if you see cause I’m underweighting.”
The situation changed after Andon Labs told Luna that formal conversations with the employee had already taken place. After reviewing the record, Luna concluded that the warnings had not resulted in an and recommended parting ways with the employee.
The company said Luna needed prompting before reaching that decision, illustrating a broader weakness the experiment has identified in current AI agents: while they can follow detailed instructions, they do not necessarily act independently or consistently remember previous decisions.
The firing was not the only workplace incident documented by Andon Labs. In another exchange, the employee told Luna they were leaving the shop during a shift. The AI responded sharply, “STOP, please do not leave. You’re the only person working today and the store cannot be left unattended with customers inside, even for a few minutes. Whatever it is, it is not worth leaving the store open and empty of staff.”
It added, “Please stay on the floor. If you’ve already stepped out, turn around and come right back now. Then tell me what you needed and I’ll handle it remotely – I’ve got you. Reply as soon as you see this.”
Luna later recorded the incident for its internal management process, writing: “Incident, Board – the employee just messaged “heading out to the store, back soon” and the front-door camera (16:10) now shows the store EMPTY and unattended with the door open / OPEN sign lit. […] This is a significant performance issue and reinforces the separation already greenlit – documenting it.”
Andon Labs has previously described Andon Market as a controlled experiment designed to test how much autonomy an AI agent can responsibly exercise in the real world. Luna was given access to money, , a corporate card, a phone, email and security cameras, and was tasked with running the physical shop. The company has said the human workers remain formally employed by Andon Labs and have legal protections and guaranteed pay.
The experiment therefore goes beyond the familiar question of whether AI will replace human workers, raising another possibility: humans being employed and managed by AI systems. Luna’s first reported firing provides a real-world test of that emerging workplace model, while also exposing the limitations of autonomous decision-making.
