I've sort of gotten on the bandwagon. I initially used AI auto complete at my previous job and liked it a lot as a better intellisense, but wouldn't have used it to write PRs -- for one I tried it a couple times and it just wasn't good enough.

My new job pushes cursor somewhat heavily and I gave it a try, it works pretty well for me although it's definitely not something I would rely on. I like being able to ask it to do something and let it go off and come back to it in a while to see how it did. For me, I think it makes it easier to start on changes by coming into something (that might be wrong and bad), but for me personally having something in a PR to start from is a nice mental hack.

If it did well enough on the initial attempt I'll try to stick with it to polish it up, but if it failed terribly I'll just write it by hand. Even when it fails it's nice to see what it did as a jumping off point. I do wish it were a bit less prone to "lying" (yada yada anthromorphization it's just tokens etc.,) though, sometimes I'll ask it to do something in a particular way (e.g., add foo to bar and make sure you X, Y, and Z) and it'll conclude (rightfully or not) that it can't do X, but then go on anyway and claim that it did X.

I wish it were easier to manage context switching in cursor though, as it is juggling IDE windows between git repo clones is a pain (this is true for everything though, so not unique to cursor). I wish I could just keep things running on a git branch and come back to them without having to manage a bunch of different clones and windows etc.,. I think this is more of a pain point with cursor since in theory it would allow you to parallelize more tasks but the tooling isn't really there.

edit: the starting point for this is probably worktrees, I remember reading about these a while ago and should probably check them out (heh) but it only solves the problem of having a bunch of clones sitting around, I'd still need to manage N windows.