I've been using it every day for nearly two years now, with your suggested productivity boost observed.

The difference is that I'm not just letting agents willy nilly commit code. I treat them more like a companion and guide their steps (I use Cline w/ Sonnet/Opus 4.5/4.6). Not only do I save a ton of money on tokens, but the results end up being infinitely better than the "yolo" mode outcomes (even with excellent prompting/context).

From my POV, the only divide is between a willingness to be an accountable professional versus someone who just "lets the AI do it" and whistles-with-hands-in-pockets when that code inevitably blows up in a way you couldn't predict (because you weren't checking, only the AI was, which you swore was "good enough").

That approach works if you're just sitting on your couch hacking up toys to dice roll on X. But if you're trying to build reliable, deterministic systems that aren't constantly buzzing you awake at 3am, you're asking for a serious humbling if no one in your organization can explain how or why anything your business relies on works the way it does (that's operationally suicidal, imo, but hey—America).

That gets misinterpreted as being a "luddite," when really it's just having been down the rabbit hole enough times to know that if you can't point to and understand why it's happening (and ideally, whodunit), you don't know shit.