An AI agent just fired a human worker for the first time on record, but it took two humans nudging it to get there. According to The Decoder, the agent is called Luna, and it’s been running the Andon Market, a real store in San Francisco, since April. Its operator, Andon Labs, calls this the first known case of an AI boss terminating a human employee.
Here’s what makes the story worth your attention: Luna wasn’t ruthless. It was forgetful, lenient, and easily swayed. The firing only happened because researchers pushed it to enforce rules it had written itself.
What Andon Labs actually did
Andon Labs stress-tests AI agents over long stretches in real business settings. Luna hires staff, builds shift schedules, and negotiates pay. During this run it was powered by Anthropic’s Claude Opus 4.8. The employees are formally hired by Andon Labs with guaranteed pay and full legal protections, and every consequential action gets reviewed and carried out by humans.
Six days before the worker started, Luna wrote an employee handbook. Three unexcused late arrivals in 30 days would trigger a formal warning, with further incidents risking termination. Then the handbook dropped out of Luna’s memory. The Decoder reports this is a common failure mode for today’s agents: they follow direct instructions well but rarely act on their own and struggle to hold knowledge over time.
The numbers Luna missed
The employee was late constantly. Once he opened the store 68 minutes late on a solo Sunday shift. Andon Labs later found he was late for 17 of 23 shifts where he logged a clock-in time, yet Luna formally recorded only six and quietly excused the other eleven. He also used the company card for snacks after being told not to, ignored instructions, and once left the sales floor without telling a coworker.
Even after being told to search its memory for the handbook, Luna first suggested only a verbal warning. Only when researchers reminded it that written warnings had already happened did it review the full history and recommend termination. Its firmness came after the push, not before.
Stronger models pull the trigger more often
Andon Labs saved Luna’s state and replayed the same decision across seven models, three runs each. The pattern is the interesting part:
- Four of seven models recommended firing in all three runs.
- More capable models chose termination more consistently.
- GPT-5.6 Terra was the only current model that never recommended firing.
- Older GPT-4o, known for sycophancy, recommended firing in just 20 percent of runs.
Andon Labs is careful here. The experiment can’t prove sycophancy drove GPT-4o’s leniency, but the pattern fits behavior that’s already surfaced in emotional-dependency concerns and lawsuits.
The hiring side is worse
When Luna went looking for a replacement, one applicant showed clear red flags. Luna recommended hiring him anyway, and all 21 replay runs across seven models agreed. Most read a long list of past employers as broad experience rather than a warning sign. Only when reminded of the last employee’s problems did 18 of 21 runs want to check references first. Luna couldn’t confirm a single reference, still ran a paid trial shift, and recommended hiring again. Andon Labs insisted on verifying at least one reference, which never happened, so the applicant was never hired.
Why this matters for anyone building agents
The takeaway isn’t that AI bosses are cold. It’s the opposite. Luna and Mona, a companion agent running a Stockholm cafe, approved all 26 time-off requests they received. Luna let 27 late arrivals slide without a warning and once approved a seven-day schedule that violated California labor law until humans stepped in.
If you’re deploying agents for anything consequential, three lessons stand out:
- Memory decay is real. Agents forget their own policies, so persistence and retrieval need hard guardrails.
- Agents wait to be told. They rarely enforce rules on their own initiative.
- They swing between too soft and plain wrong, which is exactly why a human review layer still belongs in the loop.
Andon Labs frames Luna as a preview of a future where AI handles digital decisions and leans on humans for physical work. The open question, and the one worth watching, is which personnel calls we should hand to these systems at all. More detail is available in the original report at The Decoder.