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

Synthesis: Education AI is shifting from passive chatbots to proactive agents that initiate and pursue goals. This offers personalization but risks undermining learner Learner Agency and cognitive effort and can tip into what the authors call Cognitive Surrender. The paper walks each of six pedagogical principles through what agentic initiative does to it, and proposes design responses — intentional friction, dynamic Scaffolding, human-in-the-loop oversight, and considered AI utilization.

Woollaston, Flanagan, Wijerathne & Ogata (2026, AIED HAI-Agency Workshop) review six established pedagogical principles through the lens of proactive agentic AI and articulate the central tension: the more an agent automates, the less cognitive work the learner does. Their design response — intentional friction, dynamic Scaffolding, human-in-the-loop oversight, and considered AI utilization — is a principled guardrail for the knowledge base's agentic-education literature.

The tension: automation vs. learning

Agentic systems, operating within the broader digital ecosystem of an educational application, exhibit three core behavioral traits: autonomy (functioning without continuous human intervention), proactiveness (initiating goal-directed actions), and reactiveness (adapting to changing contexts such as user input, other agents, or the digital environment). A key functional advantage is the ability to use external tools — databases, APIs, microservices — while leveraging persistent memory for cross-session continuity and multi-agent communication. While productive cognitive offloading (delegating routine mental tasks to external tools) can free working memory for higher-order reasoning, the authors warn it frequently shifts into cognitive surrender, where the learner abdicates intellectual Learner Agency and lets the AI perform the critical synthesis and analysis the student should be doing. The paper walks each of six pedagogical principles through what agentic initiative does to it:

Beyond the table, the paper flags two further risks of proactive initiative: algorithmic bias and cultural misalignment — AI agents may make flawed inferences about prior knowledge from incomplete or biased training data and often lack the nuanced local cultural competence of a human teacher, producing stereotypical or tone-deaf connections that alienate rather than engage learners.

Key Findings

• The automation–learning tension is structural. The more an agent automates a task, the less cognitive effort the learner expends — so unconstrained automation can convert productive offloading into cognitive surrender that suppresses deep processing, focus, and critical thinking.
• Six best-practice principles are each individually at risk. Prior-knowledge activation, Collaborative Learning, Problem-Based Learning, Formative Assessment, Scaffolding, and Metacognition each degrade in characteristic ways when agents take the initiative (e.g., pre-fetching bypasses retrieval practice; automated scaffolds fail to fade).
• Friction can be designed in, not avoided. The paper reframes productive struggle as a design target rather than an obstacle, echoing Desirable Difficulties theory.
• Dynamic fading is the antidote to static scaffolding. Scaffolds should adapt and be withdrawn as competence grows.
• Oversight and intentionality are prerequisites. Human-in-the-loop control over agent initiation, plus purposeful, pedagogically justified AI use, keeps automation aligned with learning rather than replacing it.

Design recommendations

1. Intentional friction — deliberately preserve productive struggle instead of maximizing smoothness (cf. Desirable Difficulties: difficulty that supports learning)
2. Dynamic scaffolding — scaffolds that adapt and fade as competence grows (cf. self-regulation)
3. Human-in-the-loop oversight — learners and educators retain control over agent initiation (Human-in-the-Loop)
4. Considered AI utilization — purposeful, pedagogically justified agent use rather than maximal automation

What this means for practice

• Instructional designers. Design friction into the agent's task loop rather than optimizing for smoothness — for example, have it ask the learner to explain how a known hobby connects to a new concept instead of drawing the connection for them.
• Instructional designers. Specify fading rules before deployment, requiring real-time analytics that distinguish temporary task completion from durable mastery so scaffolds are withdrawn rather than persisting.
• Designers. Put the teacher inside the agent's execution loop via escalation protocols, adjustable purpose and guardrails, and mid-session state interruptibility, instead of relegating educators to passive observers of chat logs.
• Designers. Apply the SAMR model feature by feature: avoid agent support that only substitutes or augments an existing task and reserve the capability for tasks it genuinely modifies or redefines.

Limitations

• This is a conceptual theoretical synthesis of six principles, not an empirical study: it reports no learner data, intervention, or outcome measures, so the implementation matrix is a design proposal rather than evidence that these strategies improve learning.
• The "learning examples" in the matrix are illustrative rather than observed classroom cases, and the application of the six principles to agentic systems is argued rather than tested.
• The recommendations — friction, dynamic fading, teacher-in-the-loop architecture, usage restraint — carry no implementation-cost, feasibility, or adoption evidence, so they remain untested design hypotheses.
• The named failure modes (cognitive surrender, learners gaming reflective checkpoints, learned helplessness) are drawn from prior literature and not measured in an agentic deployment here.

Citation

Woollaston, S., Flanagan, B., Wijerathne, I., & Ogata, H. (2026). Agentic AI and Pedagogical Best Practice: The Tension Between Automation and Learning.