ARCHIVES

Personalized education requires more than predicting whether a learner will pass a course. Predictions must support appropriate learning activities, while generated explanations must remain accurate and aligned with the curriculum. This paper proposes an AI agent-based intelligent learning system integrating learner profiling, student performance prediction, adaptive planning, retrieval-grounded content generation, and teacher review. A bounded orchestrator coordinates specialized agents through structured messages and an auditable learner-state store. Risk predictions guide the timing of support, whereas concept-level diagnostic evidence guides instructional content, this separation avoids treating course-level risk as proof of a specific misconception. The methodology combines leakage-controlled offline prediction experiments on the Open University Learning Analytics Dataset with expert evaluation of generated material and a prospective classroom study. Baselines, component ablations, calibration checks, content-quality rubrics, and learning-outcome measures are specified. The paper presents a design contribution and literature-derived findings, not an implemented or experimentally validated system. Its central contribution is a testable prediction-to-intervention workflow with explicit quality gates and human oversight. Future work will determine whether this coordinated approach improves learning and content reliability enough to justify its operational cost.