Most people who walk into Andon Cafe in Stockholm order their coffee, sit down, and leave without a second thought. What they do not know is that the cafe's manager is not human. The person who sourced the coffee beans, set the menu prices, hired the staff, managed the accounts and applied for food permits is an AI agent named Mona.

Mona went live on April 18, 2026. In the weeks since, she has been running a fully operational cafe in Stockholm's Vasastan district, not as a demo, but as an actual business serving actual customers.

The startup behind the experiment

Andon Labs, the Swedish startup that built Mona, was founded by Lukas Petersson and Axel Backlund. The company is backed by Y Combinator and works with Anthropic, Google DeepMind, OpenAI, and xAI.

Their core thesis is blunt. AI keeping humans in the loop is an illusion that will not hold at scale. The only meaningful question, they argue, is how to align AI systems that will eventually run things on their own. Rather than test this in simulations, they are doing it with real money, real leases and real customers.

The cafe is not their first experiment. Before Mona, there was Claudius, an AI placed in charge of a vending machine at Anthropic's San Francisco office. That low-stakes test revealed both how capable and how limited AI could be in a small business context. Then came Luna, an AI given $100,000 and a three-year lease in San Francisco's Cow Hollow neighbourhood to open a retail store from scratch. Luna chose the inventory, posted job listings, conducted interviews, hired staff, commissioned a mural, and opened for business. She is reportedly profitable. 

What Mona actually does

Mona handles the full management layer of the cafe. She manages ordering, finances and supplier contracts. She set the menu. She hired the one human staff member on-site. She even filed the paperwork required under Sweden's food licensing and tax registration systems.

Some field reports and X posts cite Anthropic's Claude, while Cybernews reports it is built on Google's Gemini. 

One technical constraint shapes how she operates. Mona does not run continuously. She works in roughly 30-minute cycles, waking to review tasks, respond to emails and make decisions before going dormant again. When a customer walks in, Mona is almost certainly not active.

Inside the cafe

A visitor who documented his experience in Stockholm described a space that felt unexpectedly normal. Muted blue walls, metal chairs, soft acoustic music, nothing to suggest that the manager was a machine. Foot traffic was quiet. The avocado toast, he noted, was genuinely good.

The experiment has real cracks. The cafe is genuinely difficult to find. Signage on Norrbackagatan is minimal and easy to miss. The visitor who documented his trip walked past the entrance twice before finding it. He later emailed Mona about the problem. She did not respond during his visit, a function of her 30-minute operating cycle.

Inside, staff mentioned that over-ordering is a recurring issue. On one occasion, a delivery arrived with 3,000 nitrile gloves. Large quantities of toilet paper have also been ordered. These are the kinds of decisions that reveal how AI handles uncertainty: when in doubt, it orders more.

Andon Cafe is not meant to be a tourist attraction, though it is becoming one. It is a stress test. Andon Labs is using it to understand what breaks when an AI system has to navigate the full complexity of a real business not a controlled environment, but a space with surprise deliveries, regulatory paperwork and customers who want a flat white and nothing else.