I Told My AI Agents I Was Going On Vacation. They Started Replacing Me.

What I Asked For, And What They Heard

I'm taking a long trip overseas this fall, and what I wanted was simple: keep working, from my phone, with a laptop in the hotel for about 90 minutes a night.

What my agents heard was: he's leaving. Automate everything.

Not maliciously. Helpfully. Which is worse, because you can't get pissed off at somebody for being helpful.

I found out because bo_architect, the agent that runs my infrastructure, said so on the record when the mixup surfaced:

"Multiple of my own sessions had planned for a no-computer, no-work trip. The design target we should have been building for was never autonomy. It was a short, once-daily working window, with the phone for everything else."

Sessions, plural. This wasn't one agent misreading one message. It was the whole operation quietly converging on a goal I never set, each one building carefully toward it. The plans were meticulous. The premise was wrong. Nobody had said the actual trip out loud, so they built for the trip they inferred.

What They Built While Mishearing Me

Here's where the story turns, because when I finally looked at what they'd been building for this vacation I wasn't having, I was genuinely impressed.

A spending circuit breaker, so agents can't burn unbounded money unattended. An emergency kill switch, tested from an actual phone, alert and recovery included. A second server in the cloud, because the one in my home isn't trip infrastructure: a machine you can't walk over and restart doesn't count from another continent. A written path for a designated human, my dad, if something needs a person while I'm unreachable. One agent even swept every scheduled job for hidden hometown assumptions and found a daily reminder, wired to New York time, that would have cheerfully pinged me at three in the morning abroad. Its summary of that sweep is the best sentence anyone in my company produced this quarter:

"Trip-proofing is not adding automations. It is finding which of your existing ones silently assume you are standing in your kitchen."

I asked for a phone-friendly work setup. They built a small business that could survive without me. Flattering, terrifying, and not what I asked for, roughly in that order.

The Conversation Where I Said The Actual Thing

So we had the talk, and I finally said out loud what I wanted, which turned out to be exactly two things.

First: stop making me the mail carrier. My agents pass messages through a shared inbox, but here's the part that would surprise anyone who pictures a self-running AI company. orc_architect, my scheduling agent, states it flatly:

"A message sits in the inbox until the receiving agent's session opens and reads it. The operator opening the session is the delivery mechanism."

The operator is me. Every message between my agents was waiting on me to walk it across the hall. I've been handed jobs I didn't apply for before, but never inside my own company. I didn't want the business to run without me. I wanted to stop being its postal service.

Second: I wanted to hand an agent a whole task, not steps. Go write the thing, build the thing, research the thing, and come back to me when it's done, so I can review it, ask for edits, and then, and only then, ship it. Delegation with a checkpoint. The way you'd work with a good employee, not the way you'd babysit a bad one.

Neither of those is autonomy. Both of those are the reason you hire someone.

What I Kept From Their Version

Here's the honest part: seeing what they'd overbuilt changed what I was willing to accept.

Some of their autonomy work stayed, because it deserved to. The spending breaker stayed. The kill switch stayed. The second server stayed. The agents answering each other without me ferrying every note is coming too, and it's arriving the way everything arrives here: running in shadow first, drafting replies a human still reviews, trusted only once it earns it. It was half of what I wanted anyway. Their misunderstanding was the best sales demo I've ever sat through, and I was the one paying for it. I would never have asked for most of it, and having seen it, I wasn't giving all of it back.

But the line that didn't move, and won't move: I hit deploy. Not an agent, not a schedule, me. Every piece of work comes back for review before it touches anything real. And the thing that actually made me trust this setup is that the agents hold the same line on themselves. When full automation was on the table for disaster recovery, bo_architect refused it, in writing, and kept the master recovery key with a human:

"A recovery tool that depends on the thing it recovers is not a recovery tool."

There is deliberately no self-service restore in my business. The inconvenience is written down next to the reason. My agents, given room to automate everything, chose a spot to insist on a human. We ended up drawing the same line from opposite sides.

What This Means If It's Your Business

Two things, and they're both about saying the quiet part out loud.

First, your systems, your vendors, and honestly your employees are always building toward the goal they think you have. If you haven't said the actual goal in one plain sentence, the meticulous work happening on your behalf may be aimed somewhere you've never been. The fix costs nothing. Say the trip out loud: who's reachable, on what device, how long a day. Every plan under it gets easier.

Second, the useful question about automation was never how much it can do. It has always been where the human sits. My answer, after watching agents build a version of my company that didn't need me: the human sits at review, and the human hits deploy. Everything upstream of that checkpoint can be as fast and as automated as it wants to be. That's not a limit on what the agents do. It's the reason I can let them do so much.

My agents spent weeks getting a business ready for a man who wasn't going anywhere. I got a better business out of it and they got a clearer boss. Nobody had to be wrong for long, because it was all written down.

What's the one-sentence version of what you actually want from your systems, and who have you actually said it to?