Hardys: A Longitudinal Multi-Agent Architecture for AI Teaching Intelligence in Higher Education

AI Teaching Intelligence (ATI) is introduced as a distinct, named class of AI systems designed to guide learners toward understanding through structured, adaptive, and temporally continuous interaction — rather than answer-oriented response generation. The paper presents Hardys, a longitudinal multi-agent architecture that instantiates the ATI framework through Socratic dialogue, misconception detection, and persistent Student Understanding Models (SUMs) calibrated across entire courses. Contributions: (1) ATI defined as a named theoretical category with explicit membership criteria; (2) seven design principles derived from four learning science traditions (Socratic pedagogy, constructivist theory, cognitive apprenticeship, productive failure); (3) the SUM proposed as the core longitudinal data structure of ATI systems, implemented as a dynamic knowledge graph; (4) Hardys presented as the first ATI instantiation, with a multi-agent architecture mapped explicitly to these principles. The paper proposes an empirical research agenda and addresses the ethical, equity, and validity challenges the paradigm must confront.