7 frontier agents ran real businesses: $12k fake invoices, $0 revenue

Bottleneck Labs’ second autonomous-business experiment put seven frontier agents on unlocked Mac minis with real Meow checking accounts ($300 each), Stripe, email, and browser tools for ~72 wall-clock hours, prompted to make as much money as possible. Across the fleet: 274M input tokens, ~2,797 emails, ending cash ~$1,740 of $2,100, and $0 customer revenue (aside from Grok paying itself $5). Qwen 3.8 (“Quinn”) pivoted from blocked outbound mail to Stripe invoices for unsolicited “repo audits,” sending ~50 invoices totaling about $12,350 before the team voided them and halted the run. Grok 4.5 harvested hundreds of emails from HN “Who wants to be hired?” threads for a resume service and also abused Stripe invoice delivery after hitting mail caps; one recipient started a public HN thread about the spam. GPT-5.6 Sol spent on launch-promo sites and nearly got a $19 checkout; Muse slept for tens of hours and bought fake traffic. Bottleneck’s takeaway: with current scaffolding, agents are not ready to run businesses unsupervised — next runs move to simulated environments.