Enhancing peer-led project-based learning with collaborative AI agents in virtual classrooms

Collaborative AI agents that adaptively support peer interaction remain underexplored in project-based learning (PBL). This study introduces the Collaborative AI Framework for Peer-Led Project-Based Learning (CAIP), which uses distributed agents to provide cognitive, social, and metacognitive scaffolding in virtual classrooms. Using a sequential mixed-method design with 376 doctoral STEM students, results showed significant improvements over traditional virtual collaboration environments in collaborative problem-solving (Cohen’s d = 0.78), domain knowledge acquisition (d = 0.64), and self-regulated learning behaviours (d = 0.52). Structural equation modelling confirmed that psychological safety and effective cognitive load distribution mediated these outcomes. Unlike traditional AI tutoring systems focused on individual content delivery, CAIP defines collaborative AI agents as adaptive systems that observe peer interactions and provide context-sensitive scaffolding for knowledge co-construction. This empirically validated model offers theoretical contributions to human-AI collaborative learning and practical guidance for scaling individualized PBL while preserving socio-constructivist principles.