pennware.ai

The day my AI unplugged itself

November 2026

The last big job in modernizing my home network was the switchover: retire the old router that had run the house for eight years, and let the new one take over. Claude and I had planned it carefully. The new router was built, configured and tested, and every pre-flight check passed. Early one morning I went down to the basement to move the cable.

What went wrong

The plan said that once the old router was off, Claude would finish the job: a few last steps on the new router, then checks that everything worked.

The problem was where Claude lives. The AI itself runs on Anthropic's servers, but the program I talk to it through runs on a computer in my house, and that computer reaches the internet through the old router. When I turned the old router off, Claude went silent in the middle of a step. The last thing on my screen was an error saying it couldn't reach its server.

It was a strange moment. The assistant that was supposed to finish the job had cut its own line. Nothing was harmed: the new router hadn't been changed yet, so I put the modem back on the old router and everything came back. Part of the trouble was probably the modem, too. After a change like that it needs a few minutes before it hands out a real address, and we hadn't waited.

Looking back, the mistake is obvious, and neither of us saw it. Every check in the plan tested the network. None of them asked what Claude itself depended on.

The fix

The lesson was simple: when the assistant depends on the thing being changed, it can't be the one doing the risky part. So we changed the plan.

  • Do the risky part ahead of time. Claude wrote a small script that runs on the new router by itself. It waits for the old router to go quiet, then takes over. No assistant needed.
  • Make the undo something I can do alone. A second script reverses it, and it's simple enough that I can run it without help.
  • Test both before trying again. We did.

The second try

We tried again later that same morning. This time Claude set everything up while it still had internet. It started the takeover script on the new router, and left a background job that would shut the old router down after a short delay and wake Claude up once the internet was back. Then it went offline on purpose.

The switch took about five minutes, mostly the modem starting up. My internet provider handed out a new public address, and my DNS records updated themselves within a minute. Claude came back on its own, with nobody restarting anything, and started checking.

One check failed. The service that relays my outgoing mail only accepts it from addresses I've approved, and the new address wasn't on the list. Claude found that with a test that doesn't actually send any mail. I added the address in the service's admin console, slowed down a little by the cat, who chose that moment to sit in front of the monitor, and the retest passed. Then I checked from my phone: browsing, sending mail and reading mail all worked.

What I take from this

The second try worked because it didn't need Claude during the outage. The first plan assumed the assistant would be there at the critical moment. The second plan assumed it wouldn't, and was built around that.

A careful plan can still have a blind spot about itself. The first plan was thorough and checked, and still wrong in a way that no amount of checking the network would have found. "What does the assistant need to keep working?" is now on my checklist.

Stay in the loop, and be able to act alone. I noticed the silence, and I could put things back by moving one cable. That's the job: not watching every command, but being able to step in when things go quiet.

This time it unplugged itself on purpose.