Limit it, don't inhibit it
A practical way to apply least privilege to a coding agent: a few hard walls, and plenty of room to move inside them.
// tag
#ai
9 posts
Cut along the seams
The cost of building with an agent is the length of the conversation. So the real skill is knowing where to cut a use case into pieces. I found the cut isn't where I expected: not between backend and frontend, but wherever the work first changes a promise to the outside world.
For every token it wrote, 395 were spent re-reading
The cloud bill for building the whole thing was basically nothing. The agent was a different story. And the surprising part isn't the total. It's where the tokens actually went. Part 6 of a series.
The stack is the least interesting part
The site went live for its first real match. People keep asking how it's built. But the architecture was never the interesting question. How much of it I was willing to hand to an agent was. Part 5 of a series.
I told the AI to escalate its own privileges. It refused.
Mid-task, my agent hit a permission it didn't have. I told it to grant itself the capability. It couldn't — and I've come round to thinking that's exactly right. Part 4 of a series.
The day it destroyed production
I gave my AI agent auto-approve on infrastructure changes. It saved me hours — right up until it destroyed my production host. The lesson isn't 'stop automating.' It's about where you put the tripwires. Part 3 of a series.
Give it eyes
My AI agent kept fixing a bug that wasn't fixed. It wasn't a reasoning problem — it was looking at the wrong thing. Part 2 of a series on handing an agent the whole stack.
I handed an AI agent my whole stack
Two seasons ago I shut down a fantasy football app with 500+ users because the operational toil wasn't worth it. Last week I rebuilt it on the Oracle stack — and let an AI agent run everything. Part 1 of a series.
A new home, and a small experiment
Why I moved off Wix, what I built instead, and the AI experiment tucked inside the whole thing.