I've been working on AI dev tools for a bit over a year and I don't love using AI this way either. I mostly use it for boilerplate, ideas, or to ask questions about error messages. But I've had a very open mind about it ever since I saw it oneshotting what I saw as typical Google Cloud Functions tasks (glue together some APIs, light http stuff) a year ago.

I think in the last month we've entered an inflection point with terminal "agents" and new generations of LLMs trained on their previously spotty ability to actually do the thing. It's not "there" yet and results depend on so many factors like the size of your codebase, how well-represented that kinda stuff is in its training data, etc but you really can feed these things junior-sized tickets and send them off expecting a PR to hit your tray pretty quickly.

Do I want the parts of my codebase with the tricky, important secret sauce to be written that way? Of course not, but I wouldn't give them to most other engineers either. A 5-20 person army of ~interns-newgrads is something I can leverage for a lot of the other work I do. And of course I still have to review the generated code, because it's ultimately my responsibility, but I prefer that over having to think about http response codes for my CRUD APIs. It gives me more time to focus on L7 load balancing and cluster discovery and orchestration engines.