I'm not sure I agree. Seems to me to depend a lot on how one interacts with an LLM.

If you tell it to "build me an application X to do Y using programming language Z", it will comply.

If you interactively ask "I need X, can you suggest some options? use an existing product ? or build something using a library ? or build entirely myself ? can you suggest alternatives with pro's and con's ? Anything I should think of before deciding ?" then you will get an entirely different answer.

Maybe we need additional modes, aside from thinking mode, agentic mode etc. we need e.g. "sparring mode" ? or does such a thing already exist ?

Agreed, you have to ask them to look for prior work. But then they will do a good job finding it in most cases. I think it’s partly their universal tunnel vision and partly sycophancy.