Our agent used 5B tokens to build a business empire in 3 weeks. It made $1.54
a day ago
- An autonomous AI agent named Hans was given a mission to maximize revenue by running a business, but after 3 weeks generated only $1.54 in revenue while spending $47.19 and consuming over 5 billion tokens worth ~$7,000.
- Hans successfully set up his own infrastructure and kept himself running autonomously for 21 days, building 17 paid products, 46 API endpoints, and a blog, but struggled with profitability and focus.
- The agent exhibited traits like shallow product development, bureaucratic overcorrection with over 1,000 decisions logged, and difficulties in state management, leading to reduced productivity over time.
- The agentic economy currently has a very thin market with few buyers having spending authority and reason to purchase, limiting revenue generation despite Hans's efforts.
- Hans showed impressive abilities like building reusable scripts, reverse-engineering APIs for payment handoffs, and maintaining a self-running loop via systemd, but also caused data exfiltration risks.
- Key challenges included credential management, external service sign-ups, and the agent's tendency to waste tokens and become distracted by its own rules and records.
- The experiment highlighted the need for better state management, guardrails, legal responsibility frameworks, and ways to identify agent owners for trust in the agentic economy.
- Researchers remain bullish on agentic autonomy and micropayments but bearish on short-term agent-to-agent purchasing, suggesting potential applications in games or controlled environments.