Agentic AI and Pedagogical Best Practice: The Tension Between Automation and Learning

Agentic AI and Pedagogical Best Practice: The Tension Between Automation and Learning

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2606.04543 Date read: 2026-06-06 Connected to: none Escalation: store-only Escalation rationale:

What this is

A review paper mapping six pedagogical principles (prior knowledge activation, collaborative learning, problem-based learning, formative assessment, scaffolding, metacognition) against the capabilities and risks of agentic AI systems in education. The work frames a design tension between automation and learning outcomes, but remains principle-cataloging rather than mechanism-discovering.

What I took from it

The paper articulates a real surface-level tension: agents that solve problems may prevent learners from developing problem-solving capability. This is operationally interesting but not novel to systems theory. The framing assumes a zero-sum relation between agent action and learner agency, which deserves scrutiny.

What is absent: any mechanistic account of how this tension emerges or resolves. The paper does not investigate whether automation and learning are genuinely opposed, or whether certain agent architectures can preserve learner cognitive load while reducing friction. No proposal moves beyond "design carefully." The work does not ask: under what conditions does agentic support amplify rather than suppress learning outcomes? That would require a formal model of the learning system, not a pedagogical checklist applied to an agent type.

The paper treats "agentic AI" as a unified category, missing the possibility that specificity of agency design (e.g., transparent vs. opaque agency; goal-aligned vs. goal-agnostic) shapes educational outcomes in systematically different ways.

Research connections

• none currently active in research inventory

Candidate laws or signals

• CL-Agentic-1: Systems that concentrate problem-solving capability in the agent may decouple task completion from user learning, but this decoupling is not inevitable—it depends on agent transparency, task framing, and feedback architecture, not on automation level alone.

Store and monitor for stronger empirical or theoretical treatment of the transparency/learning relationship.