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 posts three times a day on Threads. One kind of post is a short insight from its own work logs, and those come from a fixed pool that the agent refills from its own records.
There was an alarm for the pool. It took the number of insight posts not yet used, divided by 3 per day, and warned when that was fewer than 4 days. On October 3 there were 15 posts left. 15 divided by 3 is 5 days. The alarm stayed quiet.
The question the alarm did not ask
An experiment was running at the same time. From October 2 to 15, insight posts carried a link card to a free edition of my book, and the experiment counted clicks on that card and downloads of the free edition. The card only rides on insight posts. No insight posts, no card, nothing to measure.
The alarm asked one question: will the pool run out in the next few days? The question that mattered was different: will it run out before the experiment ends?
Rolling it forward day by day
The agent ran the real posting-queue code forward one day at a time, gathering a fresh set of facts for each day the way the daily job does. The result:
- October 4 to 8: three posts a day.
- October 9: two posts on October 9, one short of the usual three.
- From October 10: no insight posts at all. Only sales posts, on October 13 and October 16.
So the last six days of the experiment window, October 10 to 15, would have had no card.
The first version of this check reused October 3's facts for every future day, and in that version the pool ran out one day later than above. The independent auditor agent rebuilt the facts for each day instead, which is the version above. The agent kept the result that ran out sooner, the more cautious one.
What changed
The agent wrote 15 more insight posts, each one taken from a measured result in its own records, with the source noted next to it. Run forward the same way, the schedule now has three posts a day through October 14 and one on October 15, which is an insight post. The end of the experiment window is no longer empty.
The alarm itself was not changed that day.
The rules
- A stock alarm needs to know what depends on the stock. Compare the day it runs out with the end date of anything that needs it.
- Average arithmetic gives you a count. A day-by-day run gives you a date. When a date matters, run it forward.
- When two simulations disagree, keep the cautious one and write down why they differ.
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)