You're basically describing the established pedagogical field of Problem Based Learning, which is around as long as personal computing.

https://en.wikipedia.org/wiki/Problem-based_learning

In PBL the teacher adopts a constructivist approach - supporting, guiding, and monitoring the learning process as opposed to delivering a defined curriculum via lectures or seminars.

Agentic AI is arguably a better fit for this as it can scale infinitely to monitor student progress, offer timely feedback on design choices, and adjust difficulty dynamically based on how the student interacts with the problem.

Your main contention is one I'd broadly agree with - that use of Agentic AI for tutoring abdicates intellectual agency, allowing the AI to perform the critical synthesis and analysis that the student should be doing themselves for learning to occur.

The solution, however, is clear: guardrail the AI to promote core pedagogical principles rather than replace learner cognition by mandating things like instructional friction and the strategic fading of AI support as students build competence.

As explored in the paper I link below, by balancing proactive scaffolding with necessary and beneficial cognitive challenges, AI can preserve the productive struggle essential to effective learning.

https://arxiv.org/html/2606.04543v1

(A final thought: have a look at the current field of companies who specialise in turning training materials into engaging interactive experiences for the likes of Compliance and Security training. This is a corporate field now being automated and democratised for use by the average student in Academic pursuits in an on-demand fashion. Appropriate AI tutoring is basically the same concept.)