I would expect a naive implementation would give you a "least worse" option everyday and can't judge when it is "good enough"

Afterall, that's what most people would be when asked to make decisions for others without context.

Making the agent understand your requirements would be quite a bit of work.

Yeah that's very possible, I have never built anything with LLMs and I'm not a heavy user so I'm not sure how feasible it is.

I do think I would already get value from a least worse option every day, a sort of 'digest', so I don't have remember, and to look through results myself. I think it's a best case for LLM use for me, there is no harm at all in false negatives or positives, there are no significant stakes and I think the vagueness / unpredictability of the output is an advantage, it might find something that I had not even considered (like for my example: here is a used laptop with roughly these specs, it could also be a good home home server, something like that).