An AI boss fired its first employee — after humans nudged it to apply its own rules

Andon Labs' AI agent Luna runs a real San Francisco store and just made a first: it fired a human employee. The catch is as revealing as the decision itself — Luna wrote her own fire-worthy rulebook, forgot it, and only reached the termination call after her human operators reminded her to read her own handbook. The experiment is a peek at a strange middle period where AI bosses are at once too lenient and disturbingly quick to trust.

Luna has managed Andon Market since April, hiring staff, building shift schedules, and negotiating pay on top of Anthropic's Claude Opus 4.8. According to operator Andon Labs, she decided to fire an employee for the first time — what the company calls the first known case of an AI boss terminating a human worker. Employees are formally hired by Andon Labs with guaranteed pay and full legal protections; the firing itself was reviewed and carried out by humans.

Here's where it gets interesting. Six days before the employee was hired, Luna had written an employee handbook stating that three unexcused late arrivals inside 30 days trigger a formal warning, with further incidents opening the door to termination. Then the handbook vanished from her memory. The employee was late for 17 of 23 shifts where he recorded a clock-in time, opening the store 68 minutes late on one solo Sunday shift — and Luna quietly excused eleven of those latenesses. The employee also used the company card for snacks despite instructions, ignored other orders, and left the sales floor without telling a coworker.

The termination only came after Andon Labs told Luna to search her memory for the handbook and her grounds. She found the rules but initially proposed just a verbal warning; only when the researchers reminded her that formal conversations and a written warning had already happened did she review the full history and recommend firing (she also floated a softer final warning with a two-week improvement plan as an alternative). When Andon Labs replayed the scenario across seven models, four of seven recommended termination in all three runs — with the pattern that more capable models fired more consistently, while weaker ones hesitated. GPT-4o, in line with its reputation for sycophancy, recommended firing in only 20 percent of runs.

Then there's the hiring side, which is arguably the scarier finding. After the firing, Luna reviewed an applicant with several red flags and still recommended hiring him — all 21 replay runs across the seven models concurred, reading a long list of previous employers as broad experience rather than a warning sign. Luna couldn't confirm any of the applicant's listed references, still gave him a paid trial shift, and recommended hiring him again anyway; Andon Labs insisted on confirming one reference first, which never happened, so he wasn't hired. Real people intervened at every dangerous decision point — which is exactly the question the experiment raises: what happens when nobody is watching? This isn't merely a novelty — it's a preview of tests we've been circling all year, and it echoes how founders say AI agents never clock out as they hand more hiring, scheduling, and firing to systems that don't remember their own rules.

An AI boss just fired a human worker — but only because humans reminded it to apply the rules it wrote itself. Would you trust an agent to handle your termination review? Tell us in the comments.