What I learned building with AI coding agents

Most of what's on this site was built with AI coding agents like Claude Code and Codex. People assume that makes it easy. It makes it faster, which is not the same thing. If anything, it moved the hard part somewhere else.

The first thing I learned is that agents are happy to do the wrong thing very quickly. If you're not watching, you end up with code that looks finished and isn't. It compiles, it runs, and it's still subtly wrong in a way you only find later. That's the whole reason I built Trace. I wanted to see exactly what changed in a session and undo it in one step.

The second thing is that agents reach for filler when you don't tell them not to. Invented metrics, gradients nobody asked for, feature grids that exist to look busy. Left alone, an agent will pad a simple page into a fake startup landing page. Unslopify is just me writing those rules down once so I stop repeating them in every prompt.

The third thing is smaller but it mattered: long runs die if your laptop goes to sleep. I lost a lot of half-finished work that way before I did anything about it. ClaudeKeeper is three commands that fix exactly that.

Put together, the lesson is that the agent is fast, but you're still the one responsible. I used to think building meant typing, and now most of my time goes into reading what the agent produced, questioning it, and deciding what to throw away. The typing is the cheap part now. The judgment is the part that actually decides whether the thing is any good.