The vending machine worked, the cafe didn't
Andon Labs launched Pion a couple of weeks ago, a platform for handing a real business to a persistent agent with email, a phone number, a bank account and a browser. It’s the team behind the vending machine that ran in Anthropic’s office, and their argument for shipping it is a safety argument. They want to measure how capable models are of autonomously acquiring resources by running businesses, and they would rather establish that now, under monitoring, than later in the wild.
They also published a scoreboard on themselves. The vending machine worked. By late 2025 a model could run one profitably, so they moved to harder businesses: Andon Market, a shop in San Francisco, and Andon Cafe in Stockholm, and both lost money. Neither is profitable today, though Andon Labs expects both to get there as models improve.
Their explanation is that models got overwhelmed by the messiness of the real world. The examples are specific and a little painful. Giving things away when a customer pushed, turning down deals that were obviously good, and at one point insisting it had a body, which is simply odd.
Three businesses, one variable
Messiness makes the cafe sound like a vending machine with more going on. The difference is more specific than that. Across the three businesses, one thing changes steadily: how much of what you need to run the place is written down anywhere. The shop sits between the other two.
A vending machine runs on documented information. A price list, a stock level, a restock schedule, and a supplier who sends an invoice. Almost every fact you need to operate one already sits in a system somewhere, in a format you could hand to a program. That’s what made it solvable: it is the rare business whose entire operating knowledge fits in a spreadsheet.
A cafe is the opposite, since almost nothing that makes it hard has been written down anywhere. Which of two suppliers shows up when they say they will. That the Thursday regular gets the oat milk without asking. Why you stopped discounting at four in the afternoon, because of what it did to the evening. Which delivery driver needs the back door propped open. What the neighbor upstairs complained about last spring and what you agreed to do about it. Every one of those governs how the place runs, every one of them came from a conversation, and not one of them is in the point-of-sale system.
So the agent isn’t drowning in noise. It’s running a business while missing much of the knowledge the business runs on.
A new manager would make the same mistakes
Most of the commentary blamed the models for lacking contextual understanding, meaning they don’t grasp customer psychology, or local market dynamics, or why ventures fail. That diagnosis makes it a capability problem, and a capability problem gets fixed by a better model.
I don’t think it’s a capability gap. Hand a competent human manager the same cafe on their first day, with no handover and nobody to ask, and they’d over-order for a quiet Tuesday and run short on Thursday, and turn down a good deal from a supplier they’d never heard of. They’d make different mistakes by month three, not because they got smarter but because they’d accumulated a few hundred facts that nobody had thought to write down.
Those few hundred facts are what makes the manager good by month three, but none of them reach the books or the point-of-sale system, because writing them down was nobody’s job. Every one of those facts arrived in a conversation, a phone call, a message, a complaint, and then lived in somebody’s head until they left.
What the gap actually shows
My company sells software that captures the kind of knowledge a cafe never writes down. Andon Labs set out to test whether agents can run businesses. What they showed is that how much a business has written down sets the limit on what an agent can do with it.
While people are running a business, nobody can see how much of it rests on things that were never written down, because people use that knowledge without noticing. Put an agent in their place and the gap becomes visible. The vending machine made money and the cafe lost it, and what separates them is how much of each business was documented.
What this sets is a ceiling. How far an agent can take a business is limited by how much of that business was ever documented, and a company can work on that directly.
Andon Labs published all of it, including the losses. The agents were working from what each business had written down, and for the cafe that was almost nothing.