I'm no big fan of LLM generated code, but the fact that GP bluntly states "AI will never produce a deletion" despite this being categorically false makes it hard to take the rest of their spiel in good faith.
As a side note, I've had coworkers disappear for N days too and in that time the requirements changed (as is our business) and their lack of communication meant that their work was incompatible with the new requirements. So just because someone achieves a 10x speedup in a vacuum also isn't necessarily always a good thing.
I'd also also be wary of the risk of being an architecture-astronaut.
A declarative framework for testing may make sense in some cases, but in many cases it will just be a complicated way of scripting something you use once or twice. And when you use it you need to call up the maintainer anyway when you get lost in the yaml.
Which of course feels good for the maintainer, to feel needed.