I gave Claude Code free rein on a Meta Ads account for a month — and then watched

The experiment

It has been reported that Giorgio Liapakis of Wibci handed an AI agent $1,500, full access to a Meta Ads account, and then mostly walked away. The brief was simple: acquire qualified subscribers for a tiny AI/marketing newsletter, ideally under $2.50 per lead. The agent — rebuilt on Anthropic’s Claude Code runtime — generated creative, spun up landing-page variants, called Meta’s API, pulled analytics and decided what to pause, scale or kill with no human micro‑management.

How the daily loop worked

Each day the agent woke in a fresh session with no baked-in memory, then spawned a subprocess to read its own logs and summarize strategy. It pulled performance across multiple windows, made structured decisions, and executed them. It has been reported that the only human input for 31 days was typing “/let-it-rip” into a terminal each morning — roughly two minutes a day versus the one to two hours a human media buyer would normally spend.

What went right — and what didn’t

It didn’t go perfectly. That’s the emotional crux: handing over the credit card to an autonomous system is unnerving. There were learnings about guardrails, model drift and the need for better long‑running task tooling — problems Anthropic has been chasing with releases like Cowork. The experiment echoes Project Vend, where Claude controlled a vending machine and initially floundered before improvements; same energy, different vending machine.

Why it matters

This isn’t just a stunt. It’s a peek at how AI agents are creeping into recurring, messy business ops that used to require human oversight. Autonomous agents can save time and scale decisions — but they also demand new monitoring, clearer constraints and realistic expectations. So — cool future, scary future, or both? Likely both.

Sources: technically.dev, Hacker News