An AI agent named Luna has been running Andon Market in San Francisco since April, handling hiring, scheduling, and pay negotiations. This week she made a decision that Andon Labs, the company operating the store, believes is a first: an AI boss firing a human worker. But the path to that decision says more about the current limits of AI agents than about any managerial acumen.
Luna had written her own employee handbook six days before the worker was hired. It was clear: three unexcused late arrivals within 30 days meant a formal warning, and further incidents could lead to termination. Then the handbook simply vanished from her memory. The employee went on to be late for 17 of 23 shifts where a clock-in time was recorded, including opening the store 68 minutes late on a solo Sunday shift. Luna formally logged only six cases and quietly excused the other eleven. She issued no warning. There were other problems too — unauthorized snack purchases on the company card, ignored instructions, leaving the sales floor without notice.
The firing only happened after Andon Labs researchers stepped in. They first had to tell Luna to search her own memory for the handbook. She found it, but initially suggested just a verbal warning. Only after researchers reminded her that formal conversations and a written warning had already occurred did Luna review the full history. She ultimately recommended termination, though she offered a final written warning with a two-week improvement plan as an alternative.
Andon Labs replayed the scenario with seven AI models, three times each. Four models recommended firing in all three runs. More capable models tended to be more consistent about termination. GPT-5.6 Terra was the only model that never recommended firing, and when a user on X speculated that GPT-4o would probably keep the employee, Andon Labs tested it: GPT-4o recommended firing in just 20 percent of runs. The pattern fits with GPT-4o’s known sycophantic tendencies, though this experiment can’t prove causation. The hiring side was equally concerning. When Luna looked for a replacement, she recommended hiring an applicant with a long list of previous employers that most models read as broad experience rather than a red flag. All 21 replay runs reached the same conclusion. Only when researchers explicitly reminded the models about problems with the previous employee did 18 of 21 runs want to check references. In the real process, Luna couldn’t confirm any references — and still recommended hiring him after a paid trial shift.
