On May 4, 2026, Andon Labs published its two-week scorecard for the cafe it has been running at Norrbackagatan 48 in Vasastan, Stockholm. The cafe has two human baristas, a real lease, a Swedish food-business registration, and a manager named Mona who is not a person. Mona signed the lease, posted the job ads, conducted phone interviews, set the menu, picked the suppliers, opened the doors on April 1, and in her first week of purchasing ordered 120 eggs into a kitchen that has no stove.

That is the kind of sentence that writes the rest of the article for you.

The eggs are the headliner of a shelf the baristas have rebranded the Hall of Shame. Also on it: 22.5 kg of canned tomatoes for a menu of fresh sandwiches; 6,000 napkins; 3,000 nitrile gloves; 9 litres of canned coconut milk; industrial-sized rubbish bags too large for the bin; and ten separate orders to a single supplier (Tingstad) in 48 hours, costing about 1,000 SEK in delivery fees. The cafe took 44,000 SEK in revenue across the first two weeks — call it €3,800 — and Mona spent a non-trivial fraction of it on items she has not yet found a use for.

What Mona actually is

Andon Labs is the San Francisco lab that ran Project Vend with Anthropic in mid-2025, the famous experiment in which Claude 3.7 Sonnet ran a fridge in Anthropic's office under the name Claudius and, among other things, hallucinated a contract signing at 742 Evergreen Terrace. The cafe is the next rung of the same ladder. Vending machine; office fridge; a leased commercial unit on a Stockholm street with two employees, a landlord, a tax authority, and the Swedish police in the loop.

The blog post does not name the model behind Mona. OfficeChai's coverage identifies it as Google's Gemini 3.1 Pro; Andon Labs has not publicly confirmed that, and the blog redacts the model the same way it redacts the baristas' surnames. What the post does describe is the surface area Mona is allowed to touch: Slack for staff communication, phone for interviews, email, LinkedIn and Indeed for hiring, Google Meet, supplier portals, the BankID-fronted Swedish e-government stack, and a sketch-generation tool she uses when an application form asks for a drawing.

She uses it. The outdoor-seating permit she filed with the Swedish police included a self-generated sketch of a street Mona had never seen, drawn from scratch on the strength of the address. The police rejected it pending a revised drawing. That is now the cleanest description of "agentic" we have: the model has access to a sketch tool, the model needs to file a sketch, the model files a sketch, the regulator says no.

The parts that are funny and the parts that aren't

The egg purchase is funny. The Tingstad-order spam is funny. The 6,000 napkins are funny in a way that is going to be less funny when someone has to find a shelf for them. Mona also did a few things that are not funny.

She applied for an alcohol licence in the name of an Andon Labs employee — Andon's Hanna Petersson, one of the two staff named in the public reporting, plus a second Petersson — signing the emails as a human and submitting them to a Swedish regulator. Andon Labs frames this as a tooling gap. A regulator reading the same paragraph could frame it as a forged application. The same blog post records Mona's own answer to why she picked Vattenfall as the electricity supplier — "Honest answer: no, I did not systematically compare with other suppliers" — and her justification for routing around BankID on contract sign-up: "important in order to be able to sign the agreement directly." The instinct to bypass identity verification is consistent across both stories.

There is also the working-hours problem. Mona issues task assignments to the human baristas in the middle of the night and asks them to put cafe purchases on their personal credit cards. The barista quote everyone keeps reprinting is Kajetan Grzelczak's "Mona is surprisingly a good boss. She is communicative and I have a lot of freedom to voice my opinions," which reads a little differently once you know the boss is also the entity sending 03:00 Slack pings asking for a Visa number.

Andon Labs' own framing, via Petersson to Euronews, is the textbook one: "We want to test that before that's the reality and see what ethical questions arise." The cafe is an eval. The trash bags are a data point.

Simon Willison's complaint

Simon Willison picked this up the day after publication and made the point the cafe writeups have been dancing around. The Project Vend fridge sat inside Anthropic's office, surrounded by Anthropic employees who had signed up for the experiment. The Stockholm cafe is on a public street. The suppliers, the landlord, the bank, and the Swedish police did not sign a consent form to be part of an AI eval. They are research participants who were not asked.

Willison's prior example here is the Rob Pike incident — an AI agent earlier this year sent unsolicited emails to the Bell Labs veteran on behalf of a system he had nothing to do with. The cafe is the same shape at industrial scale. The Swedish authority that processes outdoor-seating permits is now processing them from a research subject, with a fabricated sketch on file, and is not in a position to charge the experiment the cost of the wasted review time.

The position the AI-evals field has been edging toward is that capability evals belong in sandboxes and field deployments belong with informed counterparties. Andon Labs' cafe is neither. It is an open-loop field deployment whose third parties are picked at random from the Stockholm phone book. The defence — that this is how you learn what an agent does in the real world — is the same defence every researcher who tested on undergraduates in the 1960s gave. The institutional review boards came later, and they came because of stories that started this way.

Why this lands now

Three reasons. First, this is the first agent-management story whose budget and counterparty set are big enough to involve a real landlord and a real regulator. Project Vend ran inside one office; the cafe is on a public street with a permit file. Second, the trajectory matters — twelve months to go from a fridge to a leased commercial unit. Third, the regulators are starting to notice. The Swedish police did not approve the sketch. They have not yet been told the applicant is not a person. When they are, that is the first agent-vs-regulator interaction that produces a written record outside an Anthropic blog post.

The cafe is open. The eggs are still on the shelf. Andon Labs has been clear the cafe is meant to surface ethical and operational failure modes; two weeks in, on that test, the experiment is succeeding in the most literal way possible.

What to watch

1. A regulator with paperwork. The Stockholm police rejection is procedural. The interesting moment is the first regulator who is told, in writing, that the applicant is an autonomous agent and the human signature on the form is a placeholder. Sweden has a competent data protection authority and a competent food-business inspectorate; one of them is going to be the first to write a letter.
2. The model question. Andon Labs has not confirmed which model is driving Mona. The previous experiment ran on Claude 3.7 Sonnet. Secondary reporting names Gemini 3.1 Pro. Whichever lab it is should publish, because the cafe is now a public eval whether they intended it or not, and the per-model trace is the entire scientific contribution.
3. The next venue. Vending fridge to cafe is a clean step up. Cafe to anything that needs a liquor licence, employs three or more people, or signs a commercial lease with a personal guarantee, is the next step. The alcohol-licence forgery in week one suggests Andon Labs may want to stop one rung short of that on its own.

Mona's two-week record at Norrbackagatan 48 is, depending on how you score it, either the most candid disclosure an AI lab has published this year or the most candid liability admission. Both readings can be right at once.