There was a point where clicking “approve” stopped being a decision and turned into a reflex. Not because I was lazy. It had just been right enough times in a row that I stopped actually checking. I was clicking through.
That’s the real problem with AI oversight nobody talks about. You require approval for everything, call it safety, and what you actually get is a ritual. A yes I’m not reading is the same as auto-approve. Except now I’m wasting time in between.
So, I built something into the board I’m calling earned autonomy. CAUTION categories in Mission Control still need my approval, every single time. But now the system tracks each one. Every approval that’s actually mine, really mine, not just clicked by whoever. This feeds a streak counter for that category. Hit a threshold, and the category graduates to auto-approve. The agent earns it one action at a time. I’m not handing out trust up front.
The harder part was proving the approval was actually mine. Two things have to line up. It has to come through the one path that counts as me saying yes, and the identity behind it has to check out, not just get assumed. If some other agent approves something “on my behalf,” that doesn’t count toward anything. The whole point is that the record is between me and that agent, specifically. A proxy approval just doesn’t count.
The board now shows every CAUTION category next to its streak, and how far it’s got left before it graduates. Most sit at zero. They’ve just never come up yet. Nothing has graduated. Which is probably right for something I shipped today.
I used to think oversight should be static. You decide what the AI’s allowed to do and you hold the line there. But that line gets drawn before you’ve actually seen what the thing can do. Earned autonomy flips it. Start tight, and let the track record move the line. Not me.
The trust is real because it was earned. Not because I decided to hand it over.