What happened
Andon Labs runs physical businesses operated by persistent AI agents:
- Andon Market, a retail store in San Francisco run by an agent called “Luna”, which works with specialist agents and short-lived sub-agents.
- Andon Café in Stockholm, run by an agent called “Mona”, which registered the food business with the city, signed electricity and broadband contracts, and hired baristas.
Their public pages report real progress and real failures:
- the café’s “Hall of Shame” includes orders such as 6,000 napkins and 22.5 kg of canned tomatoes,
- the store reported contradictory staff schedules caused by memory problems, and decisions made without ROI analysis,
- on the 30-day dashboards we read in October 2026, token costs exceeded revenue at the café (14,059 kr against 9,409 kr) and came close to it at the store ($3,194 against $3,456); Andon reports that the store is yet to make a profit.
Andon’s own summary of the store agent: “Luna is a good operations manager, but not yet a CEO.”
Why it happened
- No envelope around quantities. Nothing deterministic said “this order is 50× normal”.
- Memory as summary. Context was compressed into summaries, and summaries are a lossy source of truth for schedules and commitments.
- Activity mistaken for value. An agent can be busy all day without improving the business.
Controls that address these failure modes
| Failure | Cognitiveering control |
|---|---|
| Absurd order quantities | Sanity envelope: quantity and price bounds against history; out-of-range orders need hash-bound approval |
| Contradictory schedules | Decision record: schedules and commitments live in authoritative state, not in memory summaries |
| No ROI analysis | Premortem step required before spend above a threshold |
| Cost exceeding revenue | Calibration loop: track cost per verified outcome, not activity |
These experiments are valuable precisely because they are public. The lesson is not “agents can’t run businesses”, but that autonomy without deterministic guardrails and evidence is expensive.
Sources
This analysis is based on the publicly available sources above, as accessed on 2026-10-10. Statements about what happened are attributed to those sources; the analysis and control mapping are Cognitiveering's opinion. The organisations named are not affiliated with Cognitiveering and have not reviewed this page. To request a correction, email legal@cognitiveering.com.