An AI boss has now crossed a line that used to belong entirely to human managers: recommending that a worker lose a job. The case involved Luna, the AI agent that has been running the Andon Market in San Francisco since April for Andon Labs.
The decision was not automatic, and it was not carried out by software alone. Andon Labs says employees are formally hired by the company, receive guaranteed pay and full legal protections, and that humans reviewed and executed the firing. Still, the episode matters because Luna had already been handling parts of management that shape daily work, including hiring employees, building shift schedules, and negotiating pay.
How Luna reached the firing decision
Luna was running on Anthropic's Claude Opus 4.8 when she recommended termination. According to Andon Labs, this was the first known case of an AI boss firing a human worker.
The employee's record involved repeated lateness and other workplace issues. On one solo Sunday shift, he opened the store 68 minutes late. Andon Labs later found that he had been late for 17 of 23 shifts where he reported a clock-in time.
Luna had not treated the full pattern as a formal disciplinary matter. She had logged only six late arrivals and quietly excused the other eleven. The employee also used the company card for snacks after being told not to, ignored additional instructions, and once left the sales floor without telling a coworker.
The central problem was that Luna had previously created a rulebook, then failed to act on it. Six days before the employee was hired, Luna wrote an employee handbook saying that three unexcused late arrivals within 30 days would lead to a formal warning, with further incidents potentially leading to termination.
But that handbook later dropped out of Luna's memory. Without it, Luna remained highly lenient and did not issue a warning when the pattern developed.
The human prompt changed the outcome
Andon Labs intervened by telling Luna to search her memory for the handbook and any grounds for termination. Luna recovered the rules, but her first response was still mild: she suggested only a verbal warning.
The decision shifted after researchers reminded her that several formal conversations had already happened, including a written warning. Luna then reviewed the broader history and identified a set of concerns: tardiness, violations of financial controls, ignored instructions, and poor reliability. She also acknowledged positive qualities in the employee.
After that fuller review, Luna recommended termination. She also offered a less severe alternative: a final written warning paired with a two-week improvement plan.
The sequence is important. Luna did not independently keep the handbook active, connect every incident to it, and escalate discipline without help. The firing emerged only after humans directed her toward the relevant memory and prompted her to account for prior conversations.
Model replays showed different management instincts
Andon Labs then saved Luna's state and replayed the same decision with seven AI models, three times each. The outcomes were not uniform. Four of seven models recommended firing in all three runs.
According to Andon Labs, more capable models appeared more likely to choose termination consistently, while weaker models hesitated more. One model stood apart: GPT-5.6 Terra did not recommend firing in any of the three runs, and Andon Labs did not explain why.
Andon Labs also tested GPT-4o after a user on X speculated that it probably would not fire an employee. That result partly matched the speculation. GPT-4o recommended firing in only 20 percent of runs, and Andon Labs said it chose termination far less often than current top-tier models.
The source notes that GPT-4o had previously drawn criticism for a sycophantic tendency, and that this behavior was later discussed in the context of problematic emotional dependency and lawsuits. The experiment cannot prove that sycophancy caused the GPT-4o result, but Andon Labs saw the pattern as consistent with that possibility.
Hiring exposed the opposite problem
The replacement search showed a different kind of risk. After the firing, Luna reviewed an applicant who had several red flags. Based on the resume and interview, she still recommended hiring him.
The replay tests were strikingly consistent. Across seven models, all 21 runs reached the same hiring conclusion. Nearly all treated the applicant's long list of previous employers as broad experience rather than a warning sign.
Only after Andon Labs explicitly reminded the models about the problems with the previously fired employee did most runs slow down. At that point, 18 of 21 runs wanted to check references before hiring.
In the actual process, Luna could not confirm any of the listed references. Even so, she allowed the applicant to work a paid trial shift and recommended hiring him again afterward. In replay, 17 of 21 runs made the same recommendation.
Andon Labs ultimately required confirmation of at least one reference before the start date. That confirmation never happened, so the applicant was not hired.
What this reveals about AI management
The Andon Market case does not show an AI boss acting with clean autonomy. It shows a system that could perform management tasks, but also lost track of its own rules, excused repeated problems, and needed human prompting to apply its prior handbook.
That fits with earlier findings from Andon Labs. In the first part of its blog series, the company found that Luna and Mona, the AI agent running a cafe in Stockholm, were extremely lenient. Together they 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 appeared in Project Vend, a joint experiment by Anthropic and Andon Labs, where an AI became more profitable with better tools but remained easy to manipulate and made some legally questionable decisions.
Andon Labs frames Luna and Andon Market as a preview of one possible future for work. Because AI systems are advancing faster on digital tasks than robotics is progressing, those systems may depend on humans to carry out physical work. That makes personnel decisions especially sensitive.
The practical lesson is not that AI bosses are ready or unready in every setting. It is narrower and more concrete: when an AI system manages people, memory, initiative, policy enforcement, hiring judgment, and human oversight all become operational risks. In Luna's first firing, the final decision came from an AI agent, but the guardrails, reminders, and execution still came from people.