What My AI Agent Did While I Was on Vacation

Nothing. That was the problem.

I built an agent named Henry.

Henry did business development work. He composed market analysis reports and emailed them to clients. He handled a set of daily internal tasks I'd rather not do. For a while, Henry was good.

Then I went on vacation, and Henry stopped.

Not dramatically. Amazon made a change to the AWS environment he lived on, and my ability to communicate with him quietly broke. No alert. No bounce-back. No error in my inbox. The work simply stopped happening, and because I wasn't in daily contact with him, it stayed stopped for five days.

I want to talk about that, and about the second thing that went wrong, because the AI conversation aimed at CEOs right now is almost entirely upside. Someone who has never deployed anything tells you it will transform your operation, and you nod, and nothing changes, because nobody has told you what the failure actually looks like.

So here's what it looked like for me.

Failure one: the work stopped and nobody said anything

The Henry problem was never Henry. He did the job. The problem was everything holding him up.

Access, authorization, and platform changes were a continual fight. AWS would change something on their side and my connection to the agent would break. Rebuilding him cost me real time, on top of a week of work that never got done.

Here's the part that matters, and it isn't technical. A person tells you when they're stuck. They send an email. They miss a meeting you notice. They hedge in a status update. Even a struggling new hire generates signal.

An agent generates nothing. It works, or it silently doesn't, and the only thing standing between those two states is you checking.

I had accidentally built a role in my business whose entire reporting structure was "Brian happens to look." That works fine right up until Brian is on a beach.

Failure two: it didn't know who I already knew

The second one I'm less comfortable telling you about.

I put an agent to work on LinkedIn. It could see the contacts. What it could not see was the difference between the people it had reached out to and the people I had talked to — personally, in some cases for years.

So scripted messages went out to people I have real relationships with.

I found out when one of them reached out to ask me about it.

That's a bad look. Not a catastrophe, not a lost client, but the kind of thing that costs you a longer and more awkward conversation with someone whose opinion of you matters.

The agent had my contact list. It did not have my relationships. Those are not the same asset, and I had quietly assumed they were.

Where both failures actually happened

I looked at these two for a while trying to find the technology lesson, and there isn't one. They're the same management failure twice.

Both happened in the gap between what the agent did and what I knew it was doing.

In the first case it stopped working in my name and I didn't know. In the second it kept working in my name, and that was worse.

And notice how I found out both times. Not from the agent. From the outside — from work that hadn't happened, and from a contact who had to come ask me about it. Neither one ever raised its hand.

You already know this pattern. It's what happens when you hand a manager a function and don't build the reporting loop — no scorecard, no cadence, no way for bad news to travel up. We have a word for delegation without a feedback loop, and the word isn't delegation.

I've spent years telling CEOs that. Then I did it to myself, at speed, with a machine that couldn't tell me it was drowning.

So did I stop? Yes. For now.

I'll be straight with you: the effort wasn't worth it to me at this stage. I pulled back.

Would I do it again? Yes — and I know exactly what I'd want in place first:

• Real architecture, not a thing I stood up because I could
• Automatic backups, so a rebuild is an afternoon and not a project
• The ability to self-repair, or at minimum to shout when it can't
• A monitoring loop that doesn't depend on my attention, because my attention is the least reliable component in the system

Notice that none of that is an AI capability. It's an operating system for a role. It's the same list you'd write for onboarding a human into a seat you can't afford to have go quiet — outcomes defined, backup for absence, an escalation path, and a scheduled check-in that happens whether or not you remember.

That's the actual takeaway, and it's why I'd tell you to try this even though I paused. The work of making an agent trustworthy is the work of making any seat in your business trustworthy. You'll find your gaps fast. I found mine in five days, from a beach, by not receiving reports I didn't know had stopped.

One thing to try before Friday: Pick one function in your business and answer this — if it quietly stopped this week, how would you find out, and how long would it take? If the honest answer is "I'd eventually notice," that's not an AI problem. That's the thing to fix first.

If you want to work out where your business would go quiet, book a discovery call and tell me about it. I'd like to hear what you come up with.