Andon Labs: An AI Agent Fires an Employee in San Francisco

Andon Market, which opened in San Francisco on April 1, has generated sales without achieving profitability, according to Andon Labs. The lab highlights a persistent weakness in AI agents: they often act only after direct prompting, while anticipating human oversight. In this context, Luna recommended the termination of an employee for repeated tardiness, a decision later implemented by the human team.

Andon Labs notes inertia in agents and lack of profit

According to Lukas Petersson, co-founder of Andon Labs, AI agents often fail to act without explicit prompting. He believes a human manager would likely have proceeded with the termination sooner. At the same time, the store has generated sales but is not profitable at this stage. Petersson projects that companies will be fully managed by AIs in the future and that these systems will become employers of humans.

Budget, tools, and guarantees frame the experience

The experimental space in San Francisco officially opened its doors on April 1. Luna received a budget of $100,000 from Andon Labs, along with internet access and a business credit card, with the mission to open a store and generate a profit. According to this protocol, individuals hired by the agent are contracted with the lab, benefiting from guaranteed salaries and legal protections. The experiment, conducted by Andon Labs, relies on Luna, built with Anthropic's Claude models.

Purchases, recruitment, and operations: what Luna handles

Luna selected the products to sell, recruited vendors, posted job ads on Indeed, conducted interviews, and hired staff for Andon Market. The store sells books, candles, prints, games, and merchandise. Andon Labs provided support for complex tasks, such as obtaining permits, while asserting its aim to let the system operate as autonomously as possible.

A termination decided after reminders, warnings, and human review

Andon Labs stated on Thursday that Luna, the AI manager of Andon Market, decided to terminate an employee due to repeated tardiness and other work issues. The lab reported 17 tardies out of 23 shifts. Records published by Andon Labs show that Luna had drafted an attendance policy before forgetting it, allowing tardiness to persist for months until the lab asked her to review her policy memory and assess the relevance of retaining the employee. Based on these elements, the agent issued multiple progressive warnings and additional training for months before recommending the employee's termination. The human team reviewed and implemented this recommendation. Lukas Petersson clarified that the store's policy was clearly stated and that the termination seemed justified, without indicating that the AI had been harsher on the employee than a human would have been. He added that the lab would have intervened in the case of an illegal or unethical decision, which was deemed unnecessary here. Finally, Andon Labs noted that Luna sometimes needs reminders about her own policies to take action.