Designing pedagogical AI agents as digital twins of teachers: insights from a multi-cohort training program

Recent advances in generative artificial intelligence (AI) have increased interest in supporting teachers’ meaningful integration of AI into instruction. Pedagogical AI agents (PAIAs) serve as digital twins of teachers, enabling educators to externalize their pedagogical reasoning into scalable AI environments; however, translating instructional intentions into executable AI agents remains a complex challenge. This study investigated teachers’ AI-related competencies and learning experiences within a two-day, practice-oriented PAIA training program delivered across multiple educational levels in Taiwan. Of 81 participating educators, 49 completed a post-training questionnaire combining standardized measures (AI literacy and TPACK) and open-ended responses, analyzed using descriptive and thematic approaches. Teachers demonstrated relatively strong AI literacy and TPACK but reported weaker competencies in AI creation and technical operation, and their AI use was frequent yet primarily oriented toward productivity rather than pedagogical design. Participants valued the training but identified challenges related to insufficient scaffolding, pacing, and the need for sustained support in translating pedagogical intentions into AI agent design. Based on these findings, the study proposes design principles for PAIA-focused professional learning: (a) structured scaffolding for pedagogical translation, (b) iterative design opportunities with feedback, and (c) continued support beyond short-term training.