Revisiting Pedagogical Agents in the Era of Large Language Models: From Social Agency to Social Interdependence

For decades, the design of pedagogical agents (PAs) has been guided primarily by social agency theory (SAT). SAT suggests that incorporating human-like features, such as gestures, facial expressions, and voice, activates social schemas in learners, which in turn increases cognitive effort and improves learning outcomes. While this approach has been effective for traditional scripted PAs, it may not be sufficient in the era of large language models (LLMs). Unlike their predecessors, LLMs provide semantic understanding, generative capabilities, and contextual memory, fundamentally changing the nature of human-computer interaction from one-on-one, turn-based exchanges to a more collaborative social dynamic. This paper advocates for a shift in the theoretical framework from SAT to social interdependence theory (SIT) as the basis for designing LLM-driven PAs. We argue that the effectiveness of modern LLM-driven PAs is less about superficial social cues and more about the structural dynamics of collaboration: positive interdependence, promotive interaction, and individual accountability. Based on this perspective, we outline three design implications: designing for productive friction, organising interactions around shared artefacts, and positioning PAs as co-regulators of learning. We further argue that evaluation should move beyond social presence and human-likeness toward collaborative quality and shared mental model formation at the learning process level. Finally, we discuss several practical and ethical challenges. Together, these arguments reposition LLM-driven PAs not as socially appealing answer providers, but as partners in structured, accountable learning.