Notes on coding agents, so far

AI coding agents are either the future of software development or an extremely elaborate way to introduce new bugs. Probably both.

Here's what I've observed so far, handing agents real work rather than demos:

They're strongest where the work is well-defined and boring — the sort of task I'd do correctly but slowly. They're weakest exactly where I'd hoped they'd be strongest: deciding what should exist in the first place. An agent will cheerfully build the thing you asked for instead of the thing you needed, and it will do it fast enough that you don't notice for a while.

The review problem is the real problem. Generating code was never the bottleneck. Understanding code well enough to be responsible for it is the bottleneck, and that cost doesn't go away just because something else typed it.

I used to think the interesting question was how much of the work they could take over. I'm less convinced now. The interesting question seems to be what happens to my judgment when I stop doing the parts I was tempted to skip anyway.

I'm still working this out. Ask me again in six months and I may have changed my mind, which I'd count as progress.