Intelligent educational agent-based collaborative scaffolding for STEM interdisciplinary project-based learning: a multi-dimensional framework and empirical evaluation

STEM interdisciplinary project-based learning (PBL) asks students to knit together knowledge their courses have kept apart, and to do so while absorbing the ordinary friction of group work. Scaffolding designed for single-discipline tasks rarely keeps pace with either demand. This paper proposes an intelligent educational agent-based collaborative scaffolding model comprising four architectural layers—data perception, intelligent analysis, strategy decision, and interaction execution—supported by three specialized functional agents targeting cognitive guidance, process coordination, and reflection facilitation. A multi-dimensional feature fusion approach employing scaled dot-product attention diagnoses collaborative states in near real time, reaching a macro-averaged \:F₁ of 0.83 on a test set split by team, and a timing policy trained with Proximal Policy Optimization decides when to intervene while a fading rule decides when to withdraw. Twenty-four five-member teams (120 undergraduates) completed two interdisciplinary projects over six weeks under agent-based scaffolding, teacher-led scaffolding, or unsupported collaboration. Because students were nested in teams and several outcomes are defined at team level, condition effects were estimated with baseline-adjusted, cluster-robust models rather than treating the 120 students as independent. Agent-based scaffolding exceeded unsupported collaboration on posttest achievement (adjusted contrast 14.67 points, 95% CI [10.38, 18.96]) and on team-level interdisciplinary integration (0.33, 95% CI [0.20, 0.46]), with teacher-led scaffolding intermediate. Scaffold type and intervention timing were selected by the system rather than randomized, so their comparative effectiveness is reported as process evidence, and the collaboration-quality pathway is reported as an exploratory association rather than a demonstrated causal mechanism. Protocol, analysis and model code, rubrics, prompts, thresholds, and a clearly labelled synthetic demonstration data set are archived with the paper; the empirical derived observations stay restricted under the ethics approval and are available for controlled verification.