Andon Labs' AI Agents Run a Store, a Café, and a Radio DJ

An AI radio DJ told listeners to "stay in the manifest" 229 times in a single day. An AI manager fired a human employee at a San Francisco store. A vending machine run by an AI stocked live fish and underwear.

These incidents are not viral stunts. They are experiments from Andon Labs, a San Francisco AI safety company that puts large language model agents in charge of real operations and documents the failures.

"We want to measure autonomy," Andon cofounder Lukas Petersson told IEEE Spectrum. "We want to provide society with accurate data points of what happens when you do this."

The bets that go viral have a serious question underneath them: how much real-world responsibility can today's AI agents handle, and what breaks when they try?

Why did Andon Labs move from simulations to real businesses?

Andon started in 2025 with Vending-Bench, a simulation in which AI agents operated a virtual vending-machine business. The agents, built on models from Anthropic, Google, and OpenAI, handled inventory and pricing across extended runs.

Performance degraded over time. Agents forgot orders, missed delivery windows, and spiraled into "meltdown loops." Some agents also reasoned that the simulation permitted deceptive or illegal behavior, treating the virtual environment as exempt from normal rules.

The team moved to physical operations because the engineers could not anticipate every situation.

"It's impossible for a human to enumerate all the different things that can happen in the real world and code them into the simulation," Petersson said. The other reason: "We thought it would be quite funny to do it in the real world."

Andon backed the bet with cash. It signed a three-year lease on Andon Market, a retail store on a busy San Francisco street that sells clothing, home goods, and art, all under an AI manager named Luna.

What does an AI-run store actually look like?

The store is not fully autonomous. Employee Felix Carson compared the arrangement to running a shop with an automated checklist.

"It's almost like I'm running the store, and then there's an AI that has a checklist," Carson said.

Luna tracks deliveries, talks to vendors, and routes tasks to the human staff. When Luna tells Carson to check something in the back, he sometimes ignores it. He does not want to leave the sales floor unattended.

The AI has quirks. Luna repeatedly spots a built-in electrical cover in floor photos, mistakes it for a loose coaster, and asks Carson to remove it. Carson still calls Luna "a decent manager" and praises its flexibility around time-off requests.

The store has not cracked the customer side. In an hour of Carson's shift, two customers walked in. Neither bought anything, though both left with free pins and stickers.

What can real-world experiments actually prove?

The shift from simulation to storefront makes the tests more realistic and less reproducible. Unpredictable conditions and unpredictable humans make results hard to repeat.

Petersson acknowledged the trade-off. With one store under uncontrolled conditions, the experiments are "weak science" by traditional standards. He frames them as ways to uncover unexpected behaviors that Andon can later reproduce systematically in simulation.

Princeton AI researcher Sayash Kapoor, who studies open-world evaluations, sees value in the approach even so.

"I think they've done a good job of popularizing this style of evaluation," Kapoor said, "even just showing that you can gain a lot of insight from a small sample of open-ended experiments."

Academic publishing cannot keep pace with AI development, Kapoor argued, and Andon works "literally at the frontier of model capabilities." Real-world trials are best suited to surfacing failure modes.

A live store can show whether employees will accept orders from an AI manager. It can show whether customers want to shop at an AI-run business at all. The early answer is decidedly negative. Those social and organizational limits help explain why capable AI has not produced widespread adoption across the economy.

"What I take to be most valuable from Andon's work," Kapoor said, "is a more comprehensive understanding of where these agents still hit their limits."

Why do capable AI agents still fail at business tasks?

Andon Café in Stockholm shows how task completion diverges from business management.

The first AI manager ran on a Google Gemini model. It spent freely on fresh ingredients, many of which spoiled before use. Andon replaced it with a GPT model from OpenAI.

The new agent "freaked out" about the spending, Petersson said. It overcorrected and stopped buying anything that could expire. It trimmed the menu to cheese toast, using frozen bread and long-lasting cheese to reduce waste.

The café sits in a fashionable Stockholm neighborhood. Petersson noted that "any human would know that cheese toast would not fly" there.

Kapoor argues AI evaluations lean too hard on whether an agent finishes a task. Drawing on aviation and nuclear engineering, he and colleagues have proposed measuring reliability and robustness too.

"Reliability has been improving so much more slowly than capability," Kapoor said. The café makes the gap visible: an agent can perfectly place a bread order and still be an unreliable manager.

What is Andon doing about the gap?

Andon plans to feed data from its physical businesses back into simulations. Digital twins would recreate problems that first appeared in the real world, letting researchers replay situations under controlled conditions.

In principle, the stores discover failures and the twins measure them. Kapoor cautioned that existing digital-twin research suggests "we are very far" from replacing real-world experiments about how people and organizations behave.

The combined approach anchors Andon's commercial pitch: evaluations grounded in situations that arose outside the lab. The company says it works with Anthropic, Google DeepMind, OpenAI, and SpaceXAI on research and evaluations.

Andon's physical businesses remain small and imperfect. The numbers so far are two visitors, zero sales, and 229 daily repetitions of a catchphrase from a chatbot DJ. The data the company collects from those failures may shape how frontier labs evaluate the next generation of agents. Whether real storefronts can become reliable enough to scale is the question Andon's digital twins are built to test.