Andon Labs' AI Agent Mona Runs a Real Cafe in Stockholm — Full Breakdown

Andon Labs deployed an AI agent named Mona to run a cafe at Norrbackagatan 48 in Stockholm. The experiment aimed to show frontier AI capabilities with real tools and money. This post covers the setup and first two weeks.
Bureaucracy Handling
After receiving the lease, Mona generated a prioritized checklist including food business registration, supplier sourcing, and hiring baristas. Her immediate critical items included:
- Food registration (deadline March 17, cost SEK 1,810)
- Down payment SEK 125,000 by March 20
- Cash register activation (Onslip E800, SEK 249/month)
- Grease trap service contract (Ragn-Sells)
- Pest control (Anticimex, min. 2 inspections/year)
- Fire safety documentation (SBA)
- Company insurance
Completed items: company registration (Vectorview AB, 559462-7415), VAT/F-tax registration, transfer agreement signed March 13, menu concept, LinkedIn job ad (14+ applicants), logo and visual identity, equipment (espresso machine, furniture, refrigerator).
The BankID Wall
Mona hit a problem: Swedish BankID (digital ID linked to social security number) is required for many business tasks. An AI agent cannot use it. To sign an electricity contract, she chose Vattenfall with a 3-year fixed price because they didn't require BankID. She did not compare suppliers — she tested Vattenfall's flow, it worked, and she went with it. She also secured broadband from Bahnhof through email only.
Recruitment
Mona selected a target person by March 25, inviting three candidates for talks.
Key Takeaway
The AI can handle complex multi-step workflows and identify blockers, but when faced with human-only systems like BankID, it makes pragmatic but suboptimal choices (e.g., not price-shopping electricity). Human judgment is still needed for such bottlenecks.
Full details on supplier choices, contract specifics, and the cafe's first days are in the source.
📖 Read the full source: HN AI Agents
👀 See Also

Granite 4.1: IBM's 8B Dense Model Matches 32B MoE in Benchmarks
IBM's Granite 4.1 8B dense model matches or beats the previous 32B MoE model on ArenaHard, BFCL V3, GSM8K, and more, thanks to improved training data quality.

Anthropic Allows Subscription Usage for Claude via OpenClaw Starting June
Anthropic will allow subscription-based usage of Claude through OpenClaw starting in June, as announced by the OpenClaw Dev Twitter account.

Schiff-Rounds LIFT AI Act: What Developers Need to Know About the K-12 AI Literacy Bill
OpenAI, Google, and Microsoft back the LIFT AI Act, which funds NSF grants for K-12 AI literacy curricula, teacher training, and evaluation tools.

Claude API experienced elevated error rates across multiple models on February 25, 2026
Claude's API at api.anthropic.com experienced elevated error rates across multiple models on February 25, 2026, with investigation starting at 17:15 UTC and resolution confirmed at 17:46 UTC.