Andon Labs Deploys AI Agent to Run Stockholm Cafe

According to AP and PBS reporting, San Francisco-based startup Andon Labs deployed an AI agent nicknamed "Mona" to run the eponymous Andon Café in Stockholm. Per Andon Labs' blog post, the team leased space at Norrbackagatan 48 and handed Mona the lease and associated documents so the agent could begin administrative work. Multiple outlets, including PBS and Euronews, report that the agent is powered by Google's Gemini and has overseen activities from permits to hiring.

Per Andon Labs' blog, Mona parsed the lease and generated a prioritized checklist that included food business registration, finding suppliers and hiring baristas.

PBS reports the cafe has generated more than $5,700 in sales since opening in mid-April and that less than $5,000 remains from an original budget of just over $21,000, with sizable one-time setup costs cited by the experimenters. AP and other outlets document customer reactions ranging from amusement to curiosity, and PBS quotes a customer saying, "It's nice to see what happens if you push the boundary." Reporting also captures expert concern: PBS cites Emrah Karakaya, an associate professor at KTH, who warned of ethical and liability questions when AI conducts hiring or oversees food service.

Industry-pattern observations: Experiments that place agentic systems in managerial roles commonly reveal that language models excel at unstructured planning and document summarization but struggle with reliably closing external, stateful transactions that require robust error handling and accountable human oversight. For practitioners, this manifests as repeated friction points: vendor negotiations, local regulatory compliance, and edge-case customer-safety scenarios. These are not model-architecture shortcomings alone; they combine systems integration, human-in-the-loop design, and operations engineering.

Editorial analysis: The Andon Cafe experiment is a live demonstration of delegating end-to-end administrative workflows to an AI agent in a public-facing setting. It does not introduce a new modelling breakthrough, but it illustrates practical trade-offs between automation of cognitive tasks and the need for human judgment in regulated, safety-sensitive domains. Observers should treat the project as a staged probe into socio-technical boundaries rather than as evidence that current agents are ready to autonomously manage small businesses without constrained oversight.

Industry observers should monitor metrics and disclosures that indicate how failures are detected and remediated, for example, whether vendors or regulators require human signatures, how hiring decisions are documented, and how financial controls are enforced. Reporting outlets and the Andon Labs blog will be the primary sources for those operational details; Andon Labs' blog provides the most granular checklist-level evidence of what Mona attempted to complete. Ethicists and regulators quoted in mainstream coverage are likely to focus next on liability, transparency, and the legal status of AI-driven employment decisions.

Editorial analysis: For teams building agentic applications, the cafe illustrates the importance of engineering for fallbacks, audit logs, and explicit human checkpoints. Researchers and engineers should prioritize robust connectors to payment systems, formalized approval flows for regulated actions, and clear user-facing affordances that communicate when an AI is acting versus when a human is intervening.