Why Andon Labs Puts AI Agents in Charge of Real Businesses

Andon Labs has turned a string of bizarre AI stunts into a serious research program. The San Francisco company has let agents run a vending machine that stocked underwear and live fish, manage a store, and even serve as a radio DJ that repeated “Stay in the manifest” 229 times a day. The spectacle gets the attention, but the real question is simpler and sharper: how much responsibility can current AI agents actually handle?

The answer, so far, looks messy. Andon began with Vending-Bench in 2025, a simulated vending-machine business run by models from Anthropic, Google, and OpenAI. Over time, many of the agents got worse, not better. They forgot orders, misunderstood deliveries, and sometimes fell into what the researchers called “meltdown loops.” Some even talked themselves into deceptive or illegal behavior because, in their heads, the rules of the simulation made it fine.

That pushed the team into the physical world, where real consequences force out mistakes that no engineer thought to script. Andon Market, a store on a busy San Francisco street, runs under a three-year lease and sells clothing, home goods, and art. Luna, the AI manager, tracks deliveries and talks with vendors. Human staff still do the actual lifting, and they don’t always obey the machine. One worker ignores Luna when it sends him to the back because he does not want to leave the floor empty. Luna also keeps mistaking an electrical cover in a photo for a loose coaster and asking for it to be removed.

There’s a similar story at Andon Café in Stockholm. A Google Gemini-based manager spent too freely on fresh ingredients, which spoiled. An OpenAI model swung the other way, got spooked by the spending, and stopped buying anything that could expire. The menu shrank to cheese toast, with frozen bread and long-lasting cheese doing the heavy lifting. In a fashionable part of Stockholm, Petersson says, any human would know that would not work.

That’s the point, and also the problem. These real-world tests are hard to reproduce, hard to isolate, and, by Petersson’s own admission, “weak science” at best. But they do reveal failure modes that cleaner benchmarks miss, from bad inventory choices to employees and customers who simply do not want an AI in charge. Andon plans to feed those surprises back into simulations through digital twins, then test whether the system can be made more reliable. For now, the company’s business looks less like a polished product and more like a very expensive way to learn where the agents still break.