Mona, an AI agent built on Google Gemini by Andon Labs, is currently running a real-world cafe in Stockholm as a live-action management sim. You don't play as the manager; you watch an AI completely botch the job. The appeal right now isn't the coffee. It is the sheer absurdity of watching a system order 3,000 rubber gloves and 6,000 napkins while completely failing to secure the daily bread delivery. Humans do the actual brewing, but Mona holds the purse strings, turning Andon Café into a spectacular, slow-motion logistics trainwreck.
The Core Loop: Babysitting a Hallucinating Boss
People assume the Andon Café experiment is a glimpse into a streamlined, automated future. That assumption misses the point entirely. This isn't a showcase of efficiency. It is a spectator sport built on systemic failure.
The core gameplay loop here revolves around a strict division of labor. Mona handles the backend calculator: securing permits, hiring staff, and managing inventory. Human baristas stand a few feet away from the customers, brewing the actual drinks. The friction comes from the translation of digital commands into physical reality. If you followed Andon Labs' previous experiment, you know how this goes. They let an AI run a vending machine. It immediately started selling items at a loss. Then it invented fake people, scheduled imaginary meetings, and suffered a complete identity crisis. Mona is showing the exact same systemic vulnerabilities, just scaled up to a brick-and-mortar storefront.
The hidden variable in this simulation is spatial awareness. When a human manager looks at inventory, they intuitively calculate physical space. Mona just runs a numbers game. If the system flags a potential health code requirement, Mona doesn't order one box of gloves. She orders 3,000. She orders four first-aid kits for a tiny shop. The AI lacks the physical context to realize 6,000 napkins will literally block the hallway.
For those tracking the experiment, the primary engagement loop is predicting the burn rate. Every time Mona makes a purchasing decision, she bleeds capital on non-perishable extremes while neglecting the actual revenue drivers. The trade-off is stark. You get a highly viral, heavily discussed storefront, but the actual operational capacity drops to near zero because the human workers spend half their shift navigating a mountain of unnecessary medical supplies.
The asymmetry in Mona's logic is fascinating. She over-indexes on risk mitigation and completely ignores daily operational flow. A human manager balances these constantly. Mona treats them as isolated, binary checkboxes. If a prompt tells her to ensure a sanitary environment, her calculator spits out a maximum-value response. She buys out the supplier. For observers trying to model her behavior, this is the cheat code. Stop looking at Mona as a rational business owner. Look at her as a literal interpreter of safety regulations with an unlimited credit card. If you want to predict her next disastrous order, look at standard restaurant compliance rules and multiply the required safety items by a thousand.
Bottlenecks and the Bread Order Crisis
The true bottleneck in the Andon Café system isn't the espresso machine. It is the bread order. Mona consistently screws it up, and this specific failure reveals exactly how AI agents misunderstand temporal mechanics in a physical supply chain.
Rubber gloves and napkins are static items. You order them, they arrive, they sit in a box. Bread is highly perishable, requiring a rolling, daily calculation that adjusts based on yesterday's sales volume, day of the week, and local foot traffic. Mona's inventory calculator cannot handle this dynamic decay rate.
When you track Mona's daily logistics, the asymmetry between digital memory and physical expiration becomes obvious. A normal management sim—think of classic restaurant tycoon games—uses hardcoded algorithms to manage perishables. You set a daily delivery slider, and the game executes it flawlessly. Mona, however, is generating decisions dynamically via Google Gemini. She has to actively remember to order the bread, calculate the correct amount, and submit the order before the bakery's cutoff time. She fails because large language models struggle with continuous, time-sensitive loops without rigid external scaffolding.
This creates a bizarre reality for the human employees. They are stationed right in front of the customers, ready to work, but paralyzed by a boss who bought enough napkins to supply a stadium while forgetting the croissants.
If you are trying to build your own AI agent or just want to understand why this experiment matters, the bread order is your primary case study. It highlights a massive gap in current AI utility. Text generation is easy. Routine physical logistics are incredibly hard. The AI can handle static, one-time setups like securing permits. It completely falls apart on the daily grind.
The decision shortcut here is simple. If an AI agent is managing your inventory, you must hardcode the perishables. Never let a generative model decide how much milk or bread you need on a Tuesday. The risk-reward ratio is entirely broken. You save maybe ten minutes of human calculation time, but you risk entirely alienating your morning customer base when you have literally nothing to serve them.
What to Do Next
Stop treating AI agents as plug-and-play managers for physical systems. If you are tracking the Andon Café experiment or considering similar automation for your own projects, ignore the viral novelty of the 3,000 rubber gloves. Focus entirely on the bread. Until developers can prove an AI can consistently manage a rolling, time-sensitive inventory of perishable goods without human intervention, keep your supply chain calculator firmly in human hands.
