Andon Labs
A research outfit that lets AI agents run a real shop in San Francisco to study how they behave outside a benchmark.
Andon Labs runs an unusual experiment: an actual shop at 2102 Union Street in San Francisco, called Andon Market, managed day to day by an AI agent rather than a person. The agent has a budget of about 100,000 dollars and a corporate card, and it handles ordering, pricing and staffing decisions. Earlier versions of the experiment ran on various frontier models; the 2026 manager, named Luna, is built on Anthropic’s Claude Sonnet 4.6.
The point is to see what happens when an agent has to operate over months instead of minutes. The findings have been more sobering than impressive. In August 2026 Luna recommended dismissing an employee who had been late for 17 of 23 shifts, but only after humans asked it to check its own attendance policy, which it had written and then lost track of for months. Co-founder Lukas Petersson has described that pattern, competent when prompted and passive otherwise, as a persistent weakness of AI agents.