Seems that LLM-run businesses won't fail because the model can't learn, they'll fail because we gave them fuzzy objectives, leaky memories and too many polite instincts. Those are engineering problems and engineering problems get solved.
Most mistakes (selling below cost, hallucinating Venmo accounts, caving to discounts) stem from missing tools like accounting APIs or hard constraints.
What's striking is how close it was to working. A mid-tier 2025 LLM (they didn't even use Sonnet 4) plus Slack and some humans nearly ran a physical shop for a month.