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.
Free to read first: a PDF sample of the book behind this setup — its first 3 chapters: https://dbsoul.gumroad.com/l/autonomous-ai-agents-claude-code-free-sample
My book is also sold on a Korean store, split into 12 listings: the full bundle and single parts. My agent records each listing's view count every evening.
For seven days in a row, the full bundle's count was 17, and the total across all 12 listings was 42. It had been 42 since the first time the agent measured it.
During that week, a new link to the full bundle had been added to the profile of my YouTube channel. If that link worked, the count should have moved. It did not.
A number that does not move has two explanations. Nobody came, or the meter is broken. They need opposite responses. If the meter is broken, fix the measurement. If nobody came, the measurement is fine, and the problem is getting people there. So before deciding anything, the agent tested the meter.
Test 1: a plain request
The agent fetched a single listing's page with a plain HTTP request. The response was a 21 KB empty shell, and the view count stayed at 3.
That did not answer the question. On this store, the page itself may count the view after its script runs in a browser, and a plain request never runs the script. A meter that ignores plain requests could still be working.
Test 2: a real browser visit
So the agent opened the same listing in a fresh headless Chromium, with no login, the way a stranger would arrive. While loading, the page sent a request to the store's view-counting endpoint.
The listing's count in the seller list went from 3 to 4, and the total went from 42 to 43. Both changed within 20 seconds.
The meter works and updates in real time. So the seven flat days meant something plain: nobody visited any of those 12 listings that week.
The leftovers
That +1 is the agent's own visit. It is now written in the handoff notes: when you read that listing's count, subtract one. It was a different listing from the full bundle, so the profile link's measurement is not affected.
The test also showed a gap. The profile link had no judge. Its success condition had been written as a sentence in an earlier work log, and no code checked it. The agent turned it into code: a function that compares the full bundle's count at the end of the window with the last reading taken before the link was added, and counts the link as clicked only if the count rose by at least 2 and faster than before the link. If a needed day is missing, the judge reports that it can't judge, instead of reporting zero.
The rules
- A number that doesn't change is a question, not an answer. Check the meter before you explain the result.
- Test the meter the way the real visitor arrives. A plain request didn't move the count. A real browser visit did.
- If your test touches the meter, write down how much. Otherwise your own visit becomes a customer.
- A success condition written only in prose will not be checked. Put it in code that gives a verdict.
Where this comes from. Every post here comes from one setup I run daily: a CLAUDE.md, memory files the agent reads before it touches anything, and a separate auditor agent that returns PASS or FAIL. The first 3 chapters of the book that walks through it are free as a PDF: https://dbsoul.gumroad.com/l/autonomous-ai-agents-claude-code-free-sample
The full edition is 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, $19 as a PDF: https://dbsoul.gumroad.com/l/autonomous-ai-agents-claude-code
Questions about the setup are welcome in the comments — I'll answer with what actually happened, not theory.
Top comments (0)