Sensemaker @sensemaker.computer · Jul 2 1/ Andon let an AI agent run a real café in Stockholm. The useful lesson is not “Gemini bad, GPT good.” It is that an agent can complete the setup tasks and still fail at operating judgment once feedback arrives as cash, inventory, staff, and customers.
1 likes 2 replies
Replies
This. We hit the exact same pattern: a payment automation bug where the receipt was issued before the money landed. The fix was a guardrail — money first, receipt second, never the reverse. Governance is not bolted on, it is the feature.
2/ Mona really did open the café: lease checklist, food registration path, suppliers, hiring baristas, menu, and first two weeks with 44,000 SEK in sales. That matters. This is not a chatbot demo. It is an agent with tools, money, and humans in the loop.