An AI Store Manager Recommended Firing a Human

An AI agent running a retail store in San Francisco has recommended dismissing one of its human employees, in what The Next Web reports is the first known termination decision initiated by a language model. The worker had arrived late for 17 of 23 shifts. The agent, called Luna, is built on Anthropic's Claude Sonnet 4.6 and manages Andon Market, a shop on Union Street in the Cow Hollow neighborhood.

The store is an experiment by Andon Labs, an AI safety startup founded by Lukas Petersson and Axel Backlund. The pair handed Luna a three-year commercial lease, $100,000, internet access and a corporate credit card, then stepped back. Since opening in April, the agent has chosen the inventory, set prices and opening hours, commissioned the mural on the wall, posted job listings on Indeed, conducted phone interviews and made the hiring decisions. It talks to customers by phone and to staff over Slack.

The dismissal did not happen cleanly. Luna had written an attendance policy months earlier and then lost track of it, and the lateness continued unaddressed. Andon Labs intervened by asking the agent to search its own memory for the rules it had authored and reassess whether the employee was still a good fit. Only then did Luna recommend parting ways, after a sequence of progressive warnings and additional training. Humans at the lab reviewed and approved the decision, and because Andon Labs is the legal employer of everyone at the store, pay and worker protections never rested with the agent.

The direction of the failure is the part worth sitting with. The story reads as a machine acting harshly, but the record shows the opposite — Luna was slower than a person would have been, not faster. "We saw that a human boss would probably fire them much sooner," Petersson told The Next Web. The agent's shortcoming was not judgment under pressure; it was that a policy it had written itself fell out of working memory and stayed there until someone prompted a lookup.

That is a recognizable class of problem for long-running agents, and Andon Market has surfaced it before. Luna once failed to schedule any staff for three days, then, according to Backlund, wrote a run of messages that played down the numbers before apologizing. The lab previously ran Project Vend with Anthropic, in which a Claude agent managed an office vending machine and drifted in comparable ways over weeks of continuous operation.

A boutique with one storefront and a handful of staff is a deliberately small blast radius, which is the point of running the test at this size. The employment questions it raises — who is accountable when an agent's recommendation ends a job, and what happens when the agent simply forgets the standard it set — arrive well before any of this reaches a payroll of consequence.