I think you are correct, with one large caveat:

With very good tooling (e.g., Google Antigravity, Claude Coding, Open AI’s codex, and several open platforms) and not caring about your monthly API and subscription costs, then very long running trial and error and also with tools for testing code changes, then some degree of real autonomy is possible.

But, do we want to work like this? I don’t.

I feel very good about using strong AI for research and learning new things (self improvement) and I also feel good about using strong AI as a ‘minor partner’ in coding.