Key Points

• AI agent Luna, which runs a store in San Francisco for Andon Labs, decided for the first time to fire a human employee after repeated tardiness and other issues.
• Luna needed a human nudge to get there. Her self-written rulebook had dropped out of her memory, and she had largely tolerated the repeated tardiness and other problems.
• When Andon Labs replayed the scenario with seven models, more capable models tended to recommend firing more consistently.

AI agent Luna has been running a store in San Francisco since April and just fired an employee for the first time. When the scenario was replayed with different models, more capable AIs recommended termination more consistently than weaker ones.

An AI agent called Luna has been running the Andon Market in San Francisco since April. Luna has hired employees, built shift schedules, and negotiated pay. Now, according to operator Andon Labs, she's decided to fire an employee for the first time. Andon Labs says it's the first known case of an AI boss firing a human worker.

Andon Labs tests AI agents over long stretches in real business settings. Luna was running on Anthropic's Claude Opus 4.8 when she made the decision. Employees are formally hired by Andon Labs, with guaranteed pay and full legal protections. The firing was reviewed and carried out by humans.

Luna wrote her own rules, then forgot them

Six days before the employee was hired, Luna had written an employee handbook. It stated that three unexcused late arrivals within 30 days would trigger a formal warning, and further incidents could lead to termination.

Then the handbook vanished from Luna's memory. According to Andon Labs, this is a common problem with today's AI agents: they respond well to direct instructions but rarely act on their own initiative and struggle to retain knowledge over longer periods.

The employee was repeatedly late. On one occasion, he opened the store 68 minutes late on a solo Sunday shift. Luna stayed lenient and issued no warning. Andon Labs later found that the employee had been late for 17 of 23 shifts where he reported a clock-in time. Luna had only formally logged six cases and quietly excused the other eleven.

There were other problems too. The employee used the company card for snacks despite being told not to, ignored additional instructions, and once left the sales floor without telling a coworker.

The firing only happened after a human push

Andon Labs told Luna to search her memory for the handbook and any grounds for termination. She found the rules again but initially suggested only a verbal warning.

Only after the researchers reminded her that several formal conversations, including a written warning, had already taken place did Luna review the full history. She listed tardiness, violations of financial controls, ignored instructions, and poor reliability, while also acknowledging the employee's positive qualities.

She ultimately recommended termination. As an alternative, she proposed a final written warning with a two-week improvement plan. Luna needed a clear push from the outside. After that, though, her decision was firm.

Stronger models fire more readily

Andon Labs saved Luna's state and replayed the same decision with seven AI models, three times each. Four of seven recommended firing in all three runs. According to Andon Labs, a pattern seemed to emerge: more capable models chose termination more consistently, while weaker ones hesitated more often. Andon Labs did not explain why GPT-5.6 Terra was the only model that didn't recommend firing in any of the three runs.

After a user on X speculated that GPT-4o probably wouldn't fire an employee, Andon Labs ran the scenario with that model as well. The result partially proved her right: GPT-4o recommended firing in only 20 percent of runs. Andon Labs commented that the model picked termination far less often than current top-tier models.

GPT-4o had drawn criticism for its sycophantic tendency. This behavior was later discussed in the context of problematic emotional dependency and factored into lawsuits. This experiment can't prove that sycophancy drove the result, but the pattern fits.

AI models are quick to hire

After the firing, Luna looked for a replacement. One applicant brought several red flags. Based on his resume and interview, Luna still recommended hiring him. All 21 replay runs across seven models reached the same conclusion. Nearly all of them read the long list of previous employers as broad experience rather than a warning sign.

Only when Andon Labs explicitly reminded the models about the problems with the previously fired employee did 18 of 21 runs want to check references before hiring. In the actual hiring process, Luna was unable to confirm any of the listed references. She still let the applicant work a paid trial shift and recommended hiring him again afterward. 17 of 21 replay runs reached the same conclusion. Andon Labs ultimately insisted on confirming at least one reference before the start date. That never happened, so the applicant wasn't hired.

AI bosses swing between leniency and bad calls

Andon Labs had already found in the first part of its blog series that its AI bosses are extremely lenient. Luna and Mona, the AI agent running a cafe in Stockholm, approved all 26 time-off requests they received. Luna's employees were late a total of 27 times without her ever issuing a warning. Luna also approved a seven-day work schedule for one employee that, according to Andon Labs, violated California labor law, until the company stepped in.

Similar weaknesses had already surfaced during Project Vend, a joint experiment by Anthropic and Andon Labs. The AI became more profitable with better tools but remained easy to manipulate and made some legally questionable decisions.

Andon Labs sees Luna and the Andon Market as a preview of a possible future working relationship between AI and humans. Because AI is advancing faster on digital tasks than robotics is progressing, AI systems may eventually depend on humans to carry out physical work. That raises the question of which personnel decisions should be left to such systems at all.

Andon Labs