My brief experience is similar, even if conditions are perfect (you treat it like an editor, like this exactly the change you should make, flow exactly this naming and testing styles, run the tests so it's clear you didn't screw anything) it will still screw things up here and there.
People aren't impressed anymore when you slap AI to everything.
I pay for Claude and the CLI agent helps me a lot with boring stuff and that's the only AI thing I will be ever paying for. I predict price increases, and if it will exceed 50 usd, I will be out and do the boring stuff myself.
My brief experience is similar, even if conditions are perfect (you treat it like an editor, like this exactly the change you should make, flow exactly this naming and testing styles, run the tests so it's clear you didn't screw anything) it will still screw things up here and there.