What annoys me about LLMs is that they don't do fact based pushback and form cohesive implementation idea when proceeding.

Like for example if you think there's a possible problem, say that if there's a problem and suggest a workaround, the agent will change direction and implementation to that, even though there never was a problem.