Out of control

I am sitting in front of my Mac, asking Codex to delete everything we built in the past four hours. I feel a bit deflated, for sure. At the same time, I can't stop laughing.

Long story short: a while ago I built a small tool. Its pipeline depends on a third-party service to parse and download videos. I even wrote that down as a limitation, but honestly, I wrote it just for the sake of writing it. A formality. Everything has limitations, right?

Ten days later, the limitation bit me. The service stopped being open for everyone. My pipeline was dead.

So I thought: AI is a superpower, why not just build one ourselves (AI + me)?

I let Codex run. It was methodical, explaining everything as it went. And I sat there, nodding along. OK. Allow. Go ahead.

When we were almost at the finish line, one more installation to "Allow". For some reason I double-checked with Claude. Fine, it said. No risk.

Then I clicked "Deny".

An hour later I was still sitting there, trying to understand what happened. The reason was embarrassingly simple: I did not know what Codex was actually running.

I could not hold it in my head. Two agents told me everything was exactly as it should be. And still, somewhere in my brain, a warning light went on: this is out of my judgment.

Instinct, or self-protection? I can't say. Four hours of work, gone with one click. I knew what I was giving up. But I would rather waste it than keep running something I cannot judge.

And here is the part that makes me laugh. The agent was never out of control. It asked for my permission at every single step. The one who kept saying yes without understanding, for four hours straight, was me.