I mean, hallucinations are possible, especially if you have long-running agents. However, I also suspect direct instruction, simply as the alternative would take too long, be excessively expensive and the "newsworthy" result subject to chance.

Can't build a PR strategy on that.

They're trained off masses of shit with no regulation. For all we know the idea is from a mash of 2011 Reddit comments and an 2016 news article about public health in The Netherlands.

LLM control is just trial and error guardrails applied like band-aids, many of them contradicting each other.

This is what happens when technology is rushed to implementation before governments can react with regulation.