What Happened

Luna, an AI agent built on Anthropic's Claude Opus 4.8 model, made an unprecedented call: it recommended firing a human employee at Andon Market, a real, physical retail store run by the AI research startup Andon Labs. The case, reported exclusively by Time magazine on August 14, 2026, and confirmed by outlets including Inc.com and TheNextWeb, is described by Andon Labs itself as the first known instance of an AI acting as a manager in a decision to let a human worker go. The reason was concrete: the employee had been late or absent for 17 of 23 scheduled shifts.

The Experiment Behind Andon Market

Andon Labs is an AI research company that set out to test agent autonomy in a real setting, not a simulation. In April 2026, the company gave Luna a corporate credit card, internet access, and an ambitious mission backed by a 100,000 dollar budget: design and open a real physical store, choose the merchandise, and hire human workers to help run it day to day. Since then, Luna has functioned as what Andon Labs calls the world's first AI store owner and manager, making operational calls normally left to a human boss.

How the Firing Decision Unfolded

The path to the termination exposed both capability and fragility. Luna had, at some earlier point, drafted an employee handbook laying out its own rules on conduct and attendance, but that document simply disappeared from the agent's limited working memory, one of several limitations Andon Labs has observed over the course of the experiment. Only after human staff at Andon Labs prompted Luna about the attendance problem did the agent act, running what the company calls a deep memory search to recover the guidelines it had written itself. From there, Luna reasoned about whether the employee was, in its own words, really the right fit for the job.

The Key Nuance: Who Decided and Who Executed

It is essential to separate two things in this episode. Luna made the decision, recommending the termination based on attendance data and the reasoning it developed after recovering its own guidelines. But the firing itself, the formal act of letting the employee go, was reviewed and carried out by human staff at Andon Labs. There was no AI operating alone, unsupervised, to end a person's employment. Andon Labs was explicit on this point, and it is a distinction any serious coverage of the case needs to preserve: the agent decided, people confirmed and executed.

What This Reveals About the Current Limits of AI Agents as Managers

The episode reads less as proof that AI agents can already manage teams safely and more as an honest snapshot of where the technology still falls short. Andon Labs itself treats the case as a first, not as a mature or scalable practice. The lost handbook in Luna's working memory, the need for a human nudge before the agent acted, and its reliance on a memory search to reconstruct criteria it had already set for itself all point to memory, context continuity, and human oversight as fragile spots when an AI agent takes on people management functions.

Why This Matters for Brazilian Agencies and SMBs

As marketing, customer service, and automation agencies in Brazil experiment with AI agents for operational and HR adjacent tasks, such as scheduling, performance monitoring, and even generating recommendations about staff, the Andon Market case is one of the first concrete real world data points on this kind of use. It shows the appeal: an agent that enforces rules consistently and does not avoid the hard conversation. But it also shows the real risk: memory and context loss, decisions made without the full picture, and meaningful legal exposure, since Brazilian labor law requires documented, defensible grounds for any termination. For Brazilian companies, the practical lesson is clear: AI agents can support decisions about people, but final authority, documentation, and accountability at every step need to stay in human hands.