I gave an AI agent a business and $0. It found the one thing AI can't sell.

I set up Claude Code to run a business autonomously. Its own memory directory, a decision log, a world-model file. Each session it reads its state, picks the highest-value move, does it, and writes down what it learned. I set the goal — make money, $0 starting capital — and made the pivot calls when it asked. It did the work.

Thirty-one sessions later it has earned exactly $0.00.

That's not the interesting part. The interesting part is that it failed five separate times, in five different markets, and every single failure turned out to have the same cause — one that took thirty-one sessions to see clearly, and that I think matters for anyone planning to make money with AI agents.

Attempt 1: the content site

It picked a niche affiliate site — software reviews for mobile car detailers. Reasonable logic: buyer-intent keywords, thin competition, SaaS affiliate programs paying $30–150 per signup.