Purpose This pilot study explores the feasibility and user acceptance of a persistent AI agent architecture embedded in Moodle that generates misconception specific feedback and is moderated by teachers. Design/methodology/approach A four phase pilot with 100 Lebanese Grade 12 students and six physics instructors used misconception tagged Force Concept Inventory items; an OpenAI Agent Mode plugin created personalized HTML pages automatically, which teachers reviewed before release. Findings The system produced 224 pages in an average of 1.8 minutes with 99.2 percent uptime, and 83 percent were approved unchanged or after minor edits. Student surveys (N = 93) showed high acceptance across all Technology‑Acceptance‑Model constructs (means = 3.9–4.3 / 5, Cohen’s d 1), and teachers reported a 60–73 percent reduction in feedback time compared with manual preparation. Qualitative feedback highlighted diagnostic precision and the value of multimedia, while moderation logs identified curriculum alignment and scientific accuracy as priority refinement areas. Research limitations/implications This exploratory pilot focused on feasibility and user acceptance rather than learning outcomes. Future controlled studies should measure conceptual change and long-term retention to validate pedagogical effectiveness. Practical implications The framework offers institutions a replicable model for scaling personalized feedback while maintaining human oversight, with documented 60–73% time savings for instructors and high student acceptance rates. Originality/value As the first empirical demonstration of a persistent multitool AI agent inside an LMS delivering instructor verified feedback at scale, the framework is readily transferable to other STEM contexts.
Rabih Elias Kahaleh (Thu,) studied this question.