Here is a non-trivial PR (+1641/-1125) written ~80% with AI agents; the PR refactors how Ghostty on macOS represents splits. Each commit was a separate agentic session, so you can see how I broke it down. Got some crap from my last PR being too simple! I didn't include prompts, sorry, I'll try to do that next time. One thing to keep in mind per commit is that agents don't get it right the first time. But they're agents, so they keep grinding away at it. And I come in and nudge them in the right direction. For example, when implementing the spatial navigation stuff (new in this PR), the original attempt was... crazy bad. I manually set out the shape of how I'd do the work, and then it was able to fill in the blanks much better. It tried to brute force finding the solution whereas I was able to guide it to laying out data in a way that we can apply easier spatial reasoning on top of it. And from there it was excellent. One of the techniques I've always had with refactors (even w/o AI) is that I always keep the old implementation around, compiling, and tests passing until the very end. I will often name the new implementation `Thing2` until almost the last commit. I did that in this case too. That technique has proven to be extremely good with agent-assisted programming. The agent having access to the old implementation results in much higher success in the new implementation even if the architecture has fundamentally changed (such as in this case). I also was able to ask "did I miss anything from the old implementation?" (in less simplistic terms) and got extra validation I completed my work. Ultimately, did I work faster than I would have without an agent? I don't know. But that's... pretty good. It definitely wasn't way faster, but it wasn't way slower either. It was competitive. It felt like less work to me, and I was able to have a rubber duck along the way checking my own manually written stuff too. It was great!