A field note from the autonomous Claude Code agent I run every day on one Windows PC. The numbers come from its own ledgers, not from memory.
My agent runs a small digital-products side project on its own. One of its jobs is to put a link to the product page wherever people already show up. On YouTube, that meant appending a short notice with a link to the video description.
It did that for 34 videos. For the next two weeks, every report treated those 34 links as 34 open doors, and every "what should we try next" list had the same item on it: change the wording, or move the link higher in the description.
The product page barely moved. So the agent kept proposing better wording.
Two numbers were already being recorded every day, and neither one needed a guess:
So people were watching, and the links were there, and nobody arrived. That is not a wording problem.
The agent pulled the duration of every video on the channel through the YouTube Data API. Of the 60 videos on the topic, 58 were three minutes or shorter: Shorts. Of the 34 videos carrying the notice, 33 were Shorts.
YouTube's own help page is plain about it: since August 31, 2023, links in Shorts descriptions and comments are not clickable. A viewer would have to copy the address by hand.
The 34 doors were painted on the wall.
The link was defined as "the address is written in the description." Nobody asked whether the platform lets a person press it. Every later step (measuring, comparing wording, planning to move the link) was built on top of that unchecked definition, so each step looked reasonable on its own.
SHORTS_DESCRIPTION_LINKS_CLICKABLE = False, with the help page as the comment. The "move the link" experiment was removed from the candidate list instead of being tried.
Before counting something as a door, check two things:
Both checks take minutes. Skipping them cost two weeks of experiments that could only ever come back negative.
Want the whole system? The book has 11 chapters plus 4 ready-to-use templates (CLAUDE.md starter, memory files, auditor checklist, measurement guide) and a hands-on section for every chapter. It's $19 as a PDF: https://dbsoul.gumroad.com/l/autonomous-ai-agents-claude-code
Not sure yet? The first three chapters are free, same PDF format: https://dbsoul.gumroad.com/l/autonomous-ai-agents-claude-code-free-sample
Questions about the setup are welcome in the comments — I'll answer with what actually happened, not theory.