I've been trying Normal and Agent back and forth and, while I find it hard to get a definitive read on it so fast, I tend to agree that Agent sometimes tries to think way too far ahead and gets a bit eager to dive in and mess shit up (even in a good way)

On a semi-related note on Cursor: I have a theory, or more of a hunch, that I've been meaning to test:

Using Chat to craft a prompt. Tell it issues, scope, context, documentation, and have it present you with a solution, but without necessarily coding. Maybe pseudocode or steps. Ask follow-up questions, ask why he did X or Y that way, and tweak some stuff ("do it that way, not this way. You forgot to handle X Y").

Iterate until he gives you a game plan and pseudocode that makes sense.

Feed that pseudocode to Composer (Normal or Agent) like "hey this is what we're trying to do and this is the game plan so far". Observe results.

My theory is that because that game plan was generated by AI, the wording and logic is already in "AI speak", with all its quirks and ways of writing, so it will understand what you want to do with more accuracy than typing with all our human-ness.

Note: most of this is bro science coming out of my ass. Would be neat to see if results get better doing things that way.