"Just let AI write all the code" is actually reasonable if (and probably only if) almost all code of the project was generated straight from the prompt and never get checked (but it's function still might).

It's like asking AI to generate HTML pages, then just eyeball it to see if the result fits the spec, all without reading the actual code underneath.

Is it ideal and good quality? Hell no, your HTML page maybe 50MB large and comes with some extra tags that's not even in the W3C spec. But... if you observe the behavior of regular consumers and what cheap shit they got lied to paying for, you'll know that most of them don't know (or maybe even care to know) about quality.

Also there's the fact that most top LLMs can now generate code that works mostly correctly, enough to keep most consumers from complaining, all of sudden it sounded reasonable for a for-profit company to try out this route.

Again, I really don't like where this is going, but I can't ignore what I see.

A game is not a website though. The ability for subtle bugs that absolutely devastate user experience to slip through is already obviously present. Looking forward to when fixing the game breaking bug means some poor sod picking through absolute spaghetti. Oh wait they can just chuck the agents back at it and tell it to make no mistakes this time.