What my AI agents got wrong
AI agents run parts of my brand every day. This log lists every failure: what the agent was meant to do, what broke, how I caught it, what it cost and what I changed.
New entries are added when they happen. Nothing is removed.
Three articles, none scheduled
- The task
- A scheduled task writes three LinkedIn articles for the next day and schedules them, every weekday at 14:07.
- What broke
- It wrote all three articles and made their images, then scheduled none. The task ran in the cloud instead of on my computer, so it had no access to the browser signed in to LinkedIn.
- How I caught it
- I read the run report the same day.
- What it cost
- About 20 minutes of my time: I scheduled the three articles by hand in a live session.
- What I changed
- Connect the scheduled task to my computer, so it runs where the signed-in browser is.
- Status
- Fix pending. Retest on 11 November 2026 at 14:07.
The safety check stopped the task
- The task
- A scheduled agent prepared a batch of messages on a social network and was meant to send it.
- What broke
- It finished the preparation, then its own safety check refused to send. Sending cannot be undone and nobody was watching.
- How I caught it
- The run report said the batch was ready but not sent.
- What it cost
- One day's batch went out late. Nothing wrong was sent.
- What I changed
- I approve this type of action once a day in a live session. The scheduled run only prepares the batch and reminds me.
- Status
- Fixed. Running this way since 8 November 2026.
How this log works
- Only real incidents from my own projects. No client data.
- Each entry is written from the run report and checked by me before it goes live.
- Technical details that could expose accounts or systems are left out on purpose.
Want agents that fail safely?
I set up AI workflows with approval rules and final checks built in. Book a free 20-minute call.