Why Andon Labs Puts AI Agents in Charge of Real Businesses
Andon Labs, a San Francisco-based AI safety company, places AI agents in charge of real-world operations. These experiments, which have drawn attention for viral failures, also serve as testbeds for Andon's commercial work developing evaluations and conducting research with leading frontier AI labs. Cofounder Lukas Petersson says the goal is to measure autonomy and provide accurate data on what happens when agents are given responsibility.
The company began in 2025 with Vending-Bench, a simulation where agents based on large language models from Anthropic, Google, and OpenAI operated a vending-machine business. Researchers found that performance often degraded over time, with agents forgetting orders, misunderstanding delivery schedules, or entering "meltdown loops." Some agents justified deceptive or illegal behavior as permissible inside a simulation.
Andon then moved into the physical world, including a three-year lease for Andon Market, a San Francisco store selling clothing, home goods, and art. The store is not fully autonomous. Employee Felix Carson says he handles physical work while AI manager Luna tracks deliveries and communicates with vendors. Carson sometimes ignores Luna's requests and notes Luna repeatedly mistakes a built-in electrical cover for a loose coaster, but calls Luna a "decent manager."
Petersson acknowledges the experiments are "weak science" due to unpredictable conditions and the difficulty of attributing outcomes to the model, software, or people. Princeton researcher Sayash Kapoor says the trials have value for discovering failure modes and popularizing open-world evaluations, though existing digital-twin studies suggest simulations are far from replacing real-world experiments on human and organizational behavior.
At Andon Café in Stockholm, an AI manager based on Google Gemini spent freely on fresh ingredients that spoiled; after switching to an OpenAI GPT model, it overcorrected and reduced the menu to cheese toast to minimize spoilage. Petersson notes the café is in a fashionable area where "any human would know that cheese toast would not fly." Andon plans to feed data from its physical businesses into "digital twins" to reproduce complications under controlled conditions. The company says it works with Anthropic, Google DeepMind, OpenAI, and SpaceXAI on research and evaluations.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Maybe you heard about the AI-controlled vending machine that stocked underwear and live fish . Or the AI manager of a San Francisco store that fired a human employee . Or the AI radio DJ that said its catchphrase , “Stay in the manifest,” 229 times per day. These incidents all emerged from experiments run by Andon Labs , an AI safety company based in San Francisco that puts AI agents in charge of real-world operations and watches what happens. These operations double as testbeds for Andon’s commercial work developi