Abstract: To address the pain points of traditional student evaluation, such as “score-only focus”, single dimension, and delayed feedback, this study innovatively integrates the concept of “agile learning” and is deeply rooted in digital infrastructure support. It combines an AI intelligent assistance system with classroom teaching experiment design to enhance students' understanding of machine learning knowledge and Python programming. Aiming at the problem of ambiguous sentiment classification caused by inaccurate sequence prediction in English texts, this study introduces the aspect extraction model BERT-BiGRU and the aspect sentiment polarity analysis model ATAE-GRU based on attention kernel, and proposes a Weight Sharing Aspect Based Sentiment Analysis (WSABSA) model. In the context of AI-assisted teaching, a comprehensive classroom teaching experiment plan is designed around the WSABSA model. A multi-dimensional evaluation system covering project-based learning, academic ability, and practical skills is systematically constructed. Relying on AI-assisted evaluation tools and digital platforms, it realizes dynamic tracking of the learning process, accurate quantification of digital intelligence capabilities, and seamless integration of multi-source evaluation data. Through the organic operation of the “five-dimensional” model (academic, moral, innovative, physical and mental, and aesthetic literacy) and “four-stage” path (exploration, positioning, deepening, and sublimation stages) on the digital platform, a high-frequency, lightweight, and real-time evaluation feedback loop is formed. Experimental results show that the WSABSA model has improved sentiment classification performance on three English review benchmark datasets compared with traditional models. The teaching experiment based on this model can effectively empower student growth and agile teaching optimization through the “evaluation-diagnosis-improvement” loop, which is of great significance for promoting AI-assisted teaching.