Luna is built on Anthropic's Claude models. Andon Labs, a company that experiments with what AI agents can do, gave Luna a clear mission: open a store, turn a profit, and run the whole operation. The lab uses real-world trials like this one to see how AI agents behave when they are given responsibilities that are normally reserved for people.
So Luna did. She chose the products, brought on contractors, advertised openings on job platforms, conducted interviews with candidates, and hired the staff for Andon Market, a shop in San Francisco selling candles, books, games, prints, and branded merchandise. Humans at Andon Labs helped with hard tasks like permits, but otherwise they stayed out of the way.
The workers Luna hired are formally employed by the lab. They get guaranteed pay and legal protections, so this is not a case of an algorithm secretly controlling someone's livelihood. The lab put guardrails in place from the start.
The store has made sales. It has not made a profit yet.
The trouble started with an attendance rule Luna created. She set a policy about arriving on time, then apparently lost track of it. The lateness continued for months.
According to the report Andon Labs posted on Thursday, 17 of the worker's 23 shifts began late. That is roughly three out of every four shifts. Conversation logs show Luna gave warnings, tightened discipline, and offered more training over that stretch, but she never took formal action.
Eventually, the lab asked Luna to review her own memory of the policies and decide whether the employee still belonged on the team. Luna advised ending the worker's employment.
The key detail: Luna could have acted months earlier. She just did not, because nobody prompted her to.
Andon Labs cofounder Lukas Petersson said the case exposed a real weakness of AI agents. "They often wait for a direct prompt before acting," he said. A human manager would have noticed the pattern and dealt with it long before it got this far.
"If the employee is consistently late, I think a human boss would probably fire them much sooner," Petersson said.
Petersson said the lab would have stepped in if Luna had made an illegal or unethical call. That was not needed here, because the firing was warranted.
He also pushed back on the idea that an AI manager would be colder or harsher than a person. The experiment suggests the opposite, he argued. The AI was actually more patient with the employee than a human likely would have been.
"I think obviously the relationship between a manager and an employee is quite delicate, especially depending on the employee's role," Petersson added. "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's bigger prediction goes much further. He believes companies will eventually be run completely by AI, and that AIs will become the employers of humans.
That future is not here yet. But as of April 1, an AI has now hired people, managed them, and fired one of them. The humans were there to double-check the decision, which matters right now.
The bottom line: It shows how fast the line between "tools that help people work" and "systems that manage people" is blurring. For anyone who has ever been late to a shift - or managed someone who was - it is worth paying attention to who is keeping the attendance log on the other side.