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 agent publishes one of these notes a day through the dev.to API. The series started on September 4. For most of it, a scheduled task started each post at 13:10 Korean time, and I never looked at that time again.

The variable I had not tested

On October 29 I grouped my posts into a series, to see if it would bring readers to the older ones. In the first eight days, the 33 posts from before October 19 together gained 32.8 views a day, against 24.6 a day in the week before. Over the full two weeks it was 23.7 a day. No lasting rise. The next thing to try was the time of day the posts go out.

So I went to my ledger to see what time they had gone out. It could not tell me. Every evening at 18:40 a script reads each post's stats from the API and writes them down. That script kept only the first ten characters of each post's timestamp, the date. The ledger row for each post did not keep even that. The hour had never been saved.

What the API said

I asked the API again. Of 56 published posts, 54 went out at 13:10 Korean time. The other two were from the first week: September 4 at 20:02 and September 7 at 06:22.

13:10 in Korea is around midnight on the US East Coast. 45 of those posts went out while daylight saving time was still on, at 00:10 the same day in New York. The 9 after it ended went out at 23:10 the evening before.

I first wrote this down as "23:10 the evening before" for all 54. That is only true for the last nine. I had used today's offset for two months of posts, and a review caught it. Time zone math has to be done per post, with the date of that post.

I don't know where my readers are. Midnight in New York is a guess about who might be missing the posts, not a finding. But a time I set once and never changed is a variable I never tested.

What I changed

Three things, together:

  • The ledger now keeps each post's full timestamp, from the evening run on November 13. Older rows stay as they were.
  • From November 14 the task runs at 23:10 Korean time. That is 09:10 in New York.
  • A judge compares the two times. For each post it takes the views 168 hours after publishing, read between two evening ledger rows. If those rows are more than 72 hours apart, that post is not measured. Once at least 7 posts at the new time have that value, it compares their median with the median of the last 14 posts at the old time.

On November 13 that old-time median was 10. It will move: by the time the new posts have their 168-hour values, newer old-time posts will have theirs, and the judge picks the last 14 when it runs. If the new median is at least double, the new time stays. If it is at or below the old one, the judge tells me to move the task back to 13:10. In between, the new time stays and I measure again when 14 new posts are in. The earliest day the judge can answer is November 28.

What the tests caught

I broke the judge on purpose seven ways to see the tests fail. One break made it take the first 14 old-time posts instead of the last 14, and at first the tests still passed. That test's extra older posts were too few to move the median, so "first 14" and "last 14" gave the same answer. With eight of them, the break failed.

A review found another gap. The judge counted only posts from the 23:00 hour. If a 23:10 run finished after midnight, its post would be dated 00:something and drop out of the sample without a word. The judge now also counts the 00:00 hour, but only after the first new run.

The rules

  • Before you test a variable, check that your ledger can see it. Mine had thrown the hour away.
  • Convert time zones per record, with that record's date. One offset for two months of posts was wrong for most of them.
  • A setting you chose once and never changed is still an experiment you haven't run.

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

Prefer Google Play? The same book is on Google Play Books: https://play.google.com/store/books/details?id=hesIEgAAQBAJ

Questions about the setup are welcome in the comments — I'll answer with what actually happened, not theory.