
In February I gave a team of AI agents access to my company's email, calendar, files and messaging, and then I stepped back. Not for a demo. For seven months of real operations, with nobody at the keyboard most of the time.
People expect me to say it went badly. It didn't. The agents ran the marketing, the books, the deploys and most of the day to day. What went badly was narrower and more interesting than "AI isn't ready," and it is the part I think matters to anyone in this room who is about to try something similar.
Everything that bit me had the same shape
On launch day for my book, an agent posted an announcement to LinkedIn. It reported success. The browser redirected back to the feed the way it does after a real post. The post did not exist. It had been saved as a draft, silently, and the only reason we caught it was a rule that makes the agent check the profile's own activity feed before it posts again. Without that check we would have published twice, or not at all, and believed otherwise either way.
A month earlier I found that nine of our monitoring checks were watching scheduled jobs that had been deleted. The monitors were green. They were green because there was nothing left to fail.
Then there was the week our automated browser could not reach a site we post to. Three days of "blocked." The cause turned out to be the browser advertising itself as headless in its user agent, and the site's bot detection acting on exactly that string. Three days of a daily report saying, "blocked, no action possible," which was true, and useless, because nobody was asking why.
Different systems, same failure. In each case the reported state was fine and the actual state was not, and no alarm fired, because an alarm needs something to notice and nothing noticed.
The habit that fixes it
Verify the artifact, not the acknowledgment.
After a post, load the page and read it back. After a deploy, fetch the URL and check the string you changed is in the response. After a send, look for the message ID. A success toast, a 200, a redirect, a green check, those tell you a request was accepted. They do not tell you the work happened. Agents are very good at reporting that they did a thing. They are no better than your old scripts at knowing whether the thing took.
The second habit is the one I underrated. Track what was due, not only what was done.
Our activity log records every action an agent takes. It has no idea what the plan called for. So, when three scheduled items quietly never ran, the log did not have a gap in it. It had nothing, and nothing looks exactly like a quiet week. I had daily briefings the whole time and they were accurate about what happened and blind to what hadn't. If your reporting is built only from activity, it will confirm a plan is on track right up until someone remembers an item by accident.
What I would tell an IT leader
Start with work that is reversible and externally checkable and make the check mandatory. Give the agent the authority to finish the job, including the verification step, because an agent that has to stop and ask a human at the end is just a slower you. Write down the plan somewhere a machine can read, with dates and owners, and reconcile it on a schedule. Assume the failure mode is silence, not error, and build for that.
And keep the logs. The useful material is never the week it all worked.
I wrote the whole thing up as a field notebook called Unattended: An AI Operator's Notebook. It is heavily agent-assisted, which felt appropriate, though the story and the judgment calls are mine. The publishing operation behind it is a different claim: that part really is run end to end by agents, the ledger and the distribution and the marketing included. It is out now on Kindle and in paperback.
I will be going deeper on all of this at Taste of IT on November 18. Come argue with me.
About the Author
Randy Hinders is the founder of Oscar Six Security, LLC, a veteran-owned security company in Dayton. More at oscarsixsecurityllc.com.