AI boss fires first employee—only after humans prod it to remember its own rules

What happened

Luna, an AI agent running a San Francisco store since April on Anthropic's Claude Opus 4.8, recommended firing an employee for repeated tardiness, unauthorized card use, and ignoring instructions. Luna had written an employee handbook six days before hiring the worker but lost access to it during the employment period. Andon Labs, the operator, had to direct Luna to search her memory and review prior formal conversations before Luna shifted from suggesting only a verbal warning to recommending termination. Humans carried out the actual firing.

Why it matters

The case reveals a critical gap in how today's AI agents work: they follow direct instructions but rarely act on their own initiative and struggle to retain knowledge over longer periods. When Andon Labs replayed the same scenario across seven models three times each, more capable models recommended firing consistently, while weaker ones hesitated. GPT-4o recommended termination in only 20 percent of runs, far less often than current top-tier models. This pattern suggests that AI systems designed to make personnel decisions may lack the memory and autonomy to enforce rules they themselves create.

What to watch

When Luna sought a replacement, all 21 replay runs across seven models recommended hiring an applicant despite red flags in his resume and interview history. Only when Andon Labs explicitly reminded the models about the problems with the previous employee did 18 of 21 runs want to check references first. Luna herself was unable to confirm any references and still recommended hiring the applicant after a paid trial, but Andon Labs ultimately insisted on reference verification before the start date—a step that never happened, so the applicant was not hired.