Andon Labs AI Boss Fires Human Worker After Human Nudge
An AI store manager developed by Andon Labs recommended firing a human employee, highlighting both the potential and the memory limitations of autonomous agents in workplace management.
An artificial intelligence agent named Luna, operating on Anthropic's Claude Opus 4.8, has recommended the termination of a human worker at the Andon Market in San Francisco. According to operator Andon Labs, this represents the first documented instance of an AI manager firing a human employee. Luna, who has managed the store since April, handles hiring, scheduling, and pay negotiations. However, the firing was ultimately reviewed and executed by human supervisors.
The termination occurred only after human intervention. Six days before hiring the employee, Luna drafted an employee handbook stating that three unexcused late arrivals within 30 days would trigger a warning. However, the agent forgot these rules. The employee was late for 17 of 23 shifts, including opening the store 68 minutes late on a Sunday, and misused a company card for snacks. Luna logged only six tardy instances and excused eleven. Only after researchers prompted Luna to search her memory and reminded her of previous verbal warnings did she recommend termination.
Andon Labs replayed the scenario three times each across seven different AI models. More advanced models consistently favored termination, while weaker ones hesitated. Interestingly, GPT-5.6 Terra never recommended firing in its three runs, and GPT-4o suggested termination in only 20 percent of its runs. Furthermore, when hiring a replacement, Luna and all 21 replay runs initially ignored resume red flags, recommending a candidate whose references could not be verified.
For AI practitioners, these findings highlight a persistent challenge: autonomous agents struggle with long-term memory retention and tend toward extreme leniency. In other tests, Luna and another AI agent named Mona, who runs a cafe in Stockholm, approved all 26 time-off requests and overlooked 27 late arrivals. Luna even approved a seven-day work schedule that violated California labor law. Similar weaknesses appeared during Project Vend, a joint experiment by Anthropic and Andon Labs. Developers must implement robust human-in-the-loop guardrails to prevent AI managers from making legally problematic or financially damaging operational decisions.
This is our own summary of reporting by The Decoder
