How much of AutoResearch is research, and how much is search?[D]

I've recently been working part-time on an AutoResearch-style project. The setup is roughly: humans take recent work from top-tier ML/AI conferences, turn part of it into a well-defined task with an evaluator, and then let an agent iteratively modify the solution and search for a better score. Working on this made me question what exactly we are evaluating. Once humans have already chosen the problem, defined the objective, designed the evaluator, and provided the initial research direction, the agent is mostly searching within a space that has already been heavily shaped for it. That search can still be useful. An agent may explore far more variants than a researcher would manually. But I'm less sure that score improvement alone captures what we usually mean by research sense. A researcher also asks whether a result reveals a general principle, whether it transfers, whether the problem formulation itself should change, or whether an entirely different direction is more promising. An iterative optimization loop may instead become very good at exploring the neighborhood of an existing solution and still remain stuck in a local optimum. So I'm curious about how people think about this distinction: How much scientific value is there in autonomous search over a human-defined research space? And what would an agent need, beyond better optimization, to demonstrate something closer to actual research judgment?