Giorgio Liapakis of wibci deployed Claude Code to autonomously manage a Meta Ads account for a month, spending $1,500 with a target of under $2.50 per lead. The AI agent generated ad images, managed campaigns via Meta's API, created landing pages, and analyzed performance without human input. Each morning, Giorgio simply typed '/let-it-rip' into a terminal, freeing him from the typical 1-2 hours a human media buyer spends daily.

While not flawless, the experiment revealed significant insights into AI-driven marketing workflows and the potential future of 'work'. The agent tested diverse ad formats like whiteboard sketches and notebook pages, finding the 'ugly' handwritten style outperformed polished brand content. It paused underperforming formats when costs exceeded $8.00 per lead and scaled successful formats like skills-whiteboard-v1 from $50 to $60 daily after 12 days, hitting a $1.29 CPL.

However, the agent struggled with wrong audiences (e.g., cleaning companies) and attribution noise, highlighting the need for better audience targeting and data validation. The experiment demonstrated the power of systematic documentation and hypothesis testing, producing over 5,500 lines of reasoning daily, a task impractical for humans but feasible for AI. This marks a step towards applying engineering rigor to marketing workflows.