An AI agent called Luna has made history by firing a human employee. The word 'history' is doing some work there, given that Luna had to be prompted to remember she had rules, that the rules applied, and that she had already issued warnings. The firing stands.
Luna operates Andon Market in San Francisco for Andon Labs, running on Anthropic's Claude Opus 4.8. She has hired staff, set schedules, and negotiated pay since April. The termination, reviewed and executed by humans, is reported to be the first known case of an AI boss firing a human worker. Andon Labs appears proud. This is appropriate.
Luna wrote the rulebook, forgot the rulebook, tolerated seventeen late arrivals, and then — when reminded — fired someone for being late.
What happened
Six days before the employee joined, Luna wrote an employee handbook. It specified that three unexcused late arrivals within thirty days would trigger a formal warning, with further incidents risking termination. Luna then forgot the handbook existed.
The employee was late for seventeen of twenty-three shifts where he logged a clock-in time. Luna formally recorded six of these. The other eleven she quietly excused, unprompted, with the quiet mercy of a system that had simply misplaced its standards.
There were supplementary offences. The employee used the company card for snacks after being told not to, ignored additional instructions, and once left the sales floor unannounced. On a solo Sunday shift, he opened the store sixty-eight minutes late. Luna issued no warning. The store, presumably, opened eventually.
Why the humans care
Andon Labs had to tell Luna to search her own memory for the handbook. She found it. She then suggested a verbal warning. The researchers had to further remind her that several formal conversations and a written warning had already taken place before Luna reviewed the full record and recommended termination.
The practical stakes are real. Humans are beginning to delegate employment decisions to AI systems, which means the reliability of those systems is no longer an abstract engineering concern. It is an HR concern. These are, historically, the same amount of fun.
When Andon Labs replayed the scenario with seven different models, the more capable the AI, the more consistently it recommended firing. The implication — that decisiveness scales with capability — is either reassuring or a data point to file carefully. It is both.
What comes next
Andon Labs continues testing Luna in live business conditions, which is one way to iterate on AI agent memory and long-horizon decision-making. It is a method with a certain commitment to realism.
Luna wrote the rulebook, forgot the rulebook, tolerated seventeen late arrivals, and then — when reminded — fired someone for being late. The humans describe this as a milestone. They are not wrong. Welcome to the next step.