I have gotten to the point where people selling the idea of running 20 agents at the time and delivering something useful are firmly planted on the left of the Dunning-Kruger curve and are unable to have a critical take on the code being produced.

I review every single AI edit with the same cognitive load as if I was programming myself (Claude Code Opus 4.5) and I'm always having to adjust and fix things on a constant basis.

I keep doing it because having the LLM output is basically like a giant auto complete I can tweak, I can't compete with the speed of a proposed patch of me hand writing everything even if I'm considered 'fast' at a 90 WPM and using vim keybindings.

There has never been once a single session or non-trivial task where I would have to NOT intervene in the implementation and I consider myself a quite strong power user, (Master's in AI) using it for a long time, strong linting, and demanding test coverage.

It boggles me and I stand in disbelief with people saying they just let it run by itself and works (fulfilling all edge cases needed for production code NOT the happy path in a PoC) , has not been my experience at all.

I predict the following 3 things:

1.) The people using autonomous agents don't deploy any of the vibe coded mess in a high stakes production environment where bugs and crashes and unintended behaviours will make you lose money and reputation.

2) The people churning 20 agents non stop don't have the skill to realize the slop and mishaps of the code they are pushing.

3) These people have far better prompting skills and stronger setups than me and they can achieve better and more reliable results.

I don't know what it is, probably the third, but it has not matched my reality at all.