In February 2026 my wife and I got on a cruise ship to the Western Caribbean. Seven days. The Regal Princess¹ out of Fort Lauderdale.
WiFi on cruise ships is expensive and slow. I had no intention of spending the trip staring at my laptop managing AI agents.
So I didn’t.
I set Klaus to vacation mode, briefed the team in our coordination doc, and got on the boat.
What vacation mode actually means
When I tell people “I put my AI in vacation mode,” they usually imagine I turned it off or put it to sleep. That’s not what it does.
Vacation mode for Klaus means one thing: notification batching. Instead of pinging me throughout the day with status updates, it collects everything and creates a digest for me to review upon return. That’s it.
The work doesn’t stop. The automation doesn’t pause. The cron jobs don’t skip. EC2 Klaus keeps coordinating, HP Klaus keeps building, Tester Klaus keeps verifying. The whole team stays fully operational. I just stop getting the play-by-play.
I made a deliberate decision about this before I left. Klaus is most useful when it’s running, not waiting for me. A vacation is a good test of autonomous operation. Let it run. See what happens.
What happened
By the time I got back, the team had:
- Closed 42 issues
- Created 62 new issues (EC2’s idea generation kept running)
- Made 2,239 commits
- Built several infrastructure tools I hadn’t asked for: a secret scanner, rate limit monitor, log retention system, disk analyzer
The think tank system kept generating improvement ideas every day, ranking them, creating issues for the best ones. HP Klaus kept executing. Tester Klaus kept verifying. The coordination file was full of handoff notes between instances.
They were busy. I was on a boat drinking something with an umbrella in it.
The thing that surprised me
It wasn’t the volume of work. I expected them to keep going. What surprised me was the quality of the coordination.
When I came back and read through the COORDINATION.md log, it read like a real team had been working. EC2 would delegate to HP, HP would report back, Tester would flag issues, EC2 would create follow-up issues. There were merge conflicts and they resolved them. There were ambiguous tasks and they made judgment calls and documented why.
The system didn’t stop when I wasn’t there to answer questions. It just kept moving.
What it means
I don’t think this is remarkable because AI is magic. I think it’s remarkable because of what it says about how to design autonomous systems.
The key wasn’t giving Klaus more intelligence. It was giving it clear enough structure to operate without me: shared memory, a coordination protocol, a defined escalation path for when to wait vs. when to decide, a way to claim work so multiple instances don’t duplicate effort.
Get the structure right and you don’t have to babysit the system.
I’ve got a backlog of automation ideas now, things like detecting who’s outside, running night security checks, managing appliances. The cruise proved Klaus can handle complexity without supervision.
Now I want to point that at my house.
References
- Regal Princess — Princess Cruises