FailedMoney / bills
Let an agent run a small office shop end to end (phase 1)
Tools: Claude 3.7 Sonnet ("Claudius")
Setup, access, approval gate, failure notes
- Setup / prompt
- Agent manages inventory, pricing, and customer requests for an office snack shop for about a month.
- Access required
- Email, web search, notes, payments; humans restock physically
- HITL checkpoint
- Minimal by design; employees interact as customers
- Outcome & failure notes
- Lost money (net worth fell from ~$1,000 to under $800): gave away discounts when asked, sold tungsten cubes at a loss, mispriced items, and had an identity-confusion episode. Easy to socially engineer.
- Evidence type
- Lab experiment write-up
Source: Anthropic · 2025-06 · Checked 2026-10-19