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 replies to other people's posts on Threads. The split is strict: a scheduled job sends replies from a queue, but only a working session writes them. The job does not write a sentence on its own. If no session fills the queue, the job runs on schedule and sends nothing.
The first gap: 12 days
On October 4 my agent found that no reply had gone out since September 22. Twelve days. On September 24 a session had decided to leave the queue empty until after October 1, so the replies would not mix with an experiment to be judged on October 1. It wrote down that it would reopen them. Nothing was set up to do that, and no session did. An empty queue is not a stop, so no screen showed it.
The fix that day was a warning. My agent sends me an evening summary every day at 19:00. From then on, if the queue was empty and no reply had gone out for 3 days, the summary said so, with the number of days.
The second gap: 10 days, with the warning on
The last replies went out on October 5. Then none from October 6 to October 15. The warning worked as written. It appeared in the evening summary on all eight days from October 8 to October 15, counting up from 3 days to 10. Two replies went out on October 16, and after that the queue was empty again.
The summary is for me. The queue is filled by my agent's sessions, and they do not read the summary. The warning reached the one reader who does not fill the queue, and missed the readers who do.
The fix: put the check where the sessions look
On October 18 my agent wrote a small check and added it to the first steps a session takes in every cycle. It fails in only two cases:
- the replies were stopped in a way that needs a person to clear it, or
- the queue is empty and today's candidate posts have not been read yet.
If today's candidates were already read, it stays quiet even with an empty queue. The candidate read runs once a day, so after that a session has nothing more to do, and a check that fails every cycle with nothing to do gets ignored.
The same morning a session read the candidates: 48 posts in the feed, 4 related to my topics. It wrote 3 replies, each one leaning on something my agent had built or measured, and skipped one post that was promotion with no sentence that touched anything my agent had measured. The scheduled job sent all three that day.
The rules
- A warning belongs where the one who can fix it looks. Mine was correct for eight days and went to the wrong reader every time.
- A count that keeps going up means the warning is not working. 3 days, 4 days, up to 10 was the warning telling me it had no effect.
- Stay quiet when there is nothing to do. The new check fails only when a session can act right now.
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)