📄 Full text: arXiv:2606.04543 · local
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 utilisation — is a principled guardrail for the wiki's agentic-education literature.
The tension: automation vs. learning
Education AI is shifting from passive chatbots to proactive agents that initiate and pursue goals. This offers personalisation but risks undermining learner agency and cognitive effort. The paper walks each of six pedagogical principles through what agentic initiative does to it:
| Principle | Agentic-AI risk |
|---|---|
| Prior knowledge activation | Agents pre-fetching content bypass the retrieval practice that activates prior knowledge |
| Collaborative learning | Agent initiative crowds out peer negotiation and role-taking |
| Problem-based learning | Goal-directed agents resolve problems before learners grapple with them |
| Formative assessment | Agent-generated feedback pre-empts learners' own self-assessment loops |
| Scaffolding | Automated scaffolds stay static instead of being dynamically withdrawn |
| Metacognition | Agent initiation displaces the learner's own planning, monitoring, evaluation |
Design recommendations
1. Intentional friction — deliberately preserve productive struggle instead of maximising smoothness (cf. desirable-difficulties: difficulty that supports learning) 2. Dynamic scaffolding — scaffolds that adapt and fade as competence grows 3. Human-in-the-loop oversight — learners and educators retain control over agent initiation 4. Considered AI utilisation — purposeful, pedagogically justified agent use rather than maximal automation
Connections to the wiki
- Directly extends the agentic-ai corpus (agentic-workflows, tool-invariant-framework-agentic-ai) with a pedagogy-first evaluation frame
- The "teach vs. solve" concern matches measuring-llm-tutors-teach-vs-solve: agents that solve rather than scaffold fail the educational-impact test
- Intentional friction connects to desirable-difficulties and the cognitive-offloading literature
- Dynamic scaffolding aligns with scaffolding and the ZPD machinery
- Human-in-the-loop oversight echoes human-in-the-loop and the teacher-agent division of labour in agentic education workflows
- Metacognitive displacement is the same risk documented in genai-can-harm-teaching-rct-2026 (teacher-side delegation) and care-full-feedback-genai (performative reflection)
Related Pages
- agentic-ai — the technology under pedagogical scrutiny
- educational-theory — the six principles as the evaluative frame
- scaffolding — dynamic fading as a design requirement
- formative-assessment — protecting the learner's self-assessment loop
- metacognition — agency and self-regulation under agent initiative
- human-in-the-loop — oversight as a design principle
- measuring-llm-tutors-teach-vs-solve — the diagnostic counterpart to this review
- desirable-difficulties — intentional friction as a feature, not a bug
- cognitive-offloading — the automation-learning trade-off mechanism
- agentic-workflows — agentic education practice
- genai-can-harm-teaching-rct-2026 — empirical evidence of automation displacing pedagogy
Sources
- Woollaston, S., Flanagan, B., Wijerathne, I., & Ogata, H. (2026). Agentic AI and Pedagogical Best Practice: The Tension Between Automation and Learning. arXiv:2606.04543