Running AI agents unattended: 3 failure modes from one week

Description and use

Three real incidents from logs this week show how GUI automations, account-sign-in checks, and agent self-verification can fail when running AI agents unattended for a solo developer's Korean information site. A shared Chrome window caused a GUI collision despite a lock file; an upload failed because the wrong channel was signed in; the video agent missed duplicates and miscounted. The author adopted pre-publish account checks, hand-publishing the first item before letting loops run, and spot-checking agent reports against files.

Problem

Three real incidents from logs this week show how GUI automations, account-sign-in checks, and agent self-verification can fail when running AI agents (Claude Code) unattended for a solo developer's Korean information site.

Outcome

Three failure modes were observed: GUI automation collisions despite a lock file, an upload failing because the wrong channel was signed in, and the video agent missing duplicates and miscounting. The author adopted pre-publish account checks, hand-publishing the first item before letting loops run, and spot-checking agent reports against files.

Workflow

• Paste a draft into a page while another loop navigated the same window elsewhere, causing input to land on the wrong screen.
• Check whether the lock prevented the overlap; nothing had been posted but the cause remains unknown.
• Read the current signed-in account before publishing and refuse to run on a mismatch.
• Publish only the first item by hand before letting a loop run.
• Spot-check the video agent's topic existence check against the actual files, since seven of eighteen videos overlapped with existing ones.
• Spot-check the video agent's count report, since it reported 145 left when the real number was 176.

Evidence

Some of my publishers drive a real Chrome window with clicks and paste, so they all share one browser. I use a lock file so they take turns. Today I pasted a draft into a page while another loop navigated the same window elsewhere, so my input landed on the wrong screen.

It was before any submit button, and I checked afterwards that nothing had been posted, but I still don't know why the lock didn't prevent the overlap.

The fix I'm adopting: read the current account first and refuse to run on a mismatch, and publish only the first item by hand before letting a loop run.

The video agent checked whether a topic already existed, but only matched one way of writing the identifier in the code. Seven of eighteen videos overlapped with existing ones.

The same agent also reported '145 left' when the real number was 176.