The accelerating integration of Artificial Intelligence (AI) into higher education has enabled scalable academic support systems, yet persistent concerns regarding hallucination, contextual drift, and limited accountability continue to constrain the dependable deployment of Large Language Model (LLM)-based teaching assistants. In response to these limitations, this study introduces an Agentic Retrieval-Augmented Generation (Agentic RAG) framework that redefines the operational paradigm of AI teaching assistants from reactive information providers to predictive, self-regulating academic collaborators. The proposed system leverages a hybrid retrieval mechanism combining dense vector similarity search, BM25-based sparse indexing, and curriculum-aware knowledge graph traversal to ensure semantically grounded response generation and consistent instructional reasoning. A central innovation of the framework lies in its predictive reasoning layer, where student interaction dynamics are modeled as a probabilistic learning state estimation problem. Formally, the likelihood of a knowledge deficiency is expressed as \begin{equation} P(Gap_t \mid \mathbf{x}) = \sigma(\mathbf{w}^{T}\mathbf{x} + b) \end{equation} where $\mathbf{x}$ denotes behavioral indicators such as repeated query frequency, task completion latency, and historical error rates, and $\sigma(\cdot)$ represents the logistic activation function. This formulation enables early detection of learning instability and supports pre-emptive scaffolding interventions aligned with adaptive instructional strategies. To strengthen system reliability, a trust validation module computes a composite response confidence metric defined as \begin{equation} T = \alpha A + \beta C + \gamma R \end{equation} where $A$ denotes factual accuracy, $C$ citation consistency, and $R$ contextual relevance derived from cross-encoder re-ranking scores. The architecture was evaluated using a controlled academic simulation environment constructed from multi-domain university course datasets, including structured lecture repositories, assessment records, and anonymized student interaction logs. Experimental results demonstrate that the proposed Agentic RAG system achieved a response accuracy of \textbf{97\%}, outperforming traditional Retrieval-Augmented Generation systems that recorded \textbf{92\%} accuracy and significantly surpassing standalone language models with \textbf{72\%} accuracy. In parallel, the hallucination rate was reduced to \textbf{1.5\%}, compared to \textbf{4\%} in conventional RAG architectures and \textbf{18\%} in baseline LLM implementations, indicating a substantial improvement in response reliability and factual grounding. Furthermore, the system attained a trust evaluation score of \textbf{9.3}, exceeding the comparative trust scores of \textbf{8.7} and \textbf{5.2} observed in traditional RAG and standalone LLM systems, respectively. These quantitative outcomes confirm that the integration of agentic reasoning and trust-aware validation mechanisms contributes to stable inference performance while maintaining acceptable response latency under real-time academic workloads. Therefore, this work contributes a rigorously engineered and empirically validated framework for autonomous academic support systems, demonstrating that predictive retrieval orchestration and reliability-aware reasoning can substantially enhance the trustworthiness, scalability, and pedagogical effectiveness of next-generation AI teaching assistants.

Keywords: Agentic Artificial Intelligence, Retrieval-Augmented Generation (RAG), Predictive Learning Analytics, Trustworthy AI, Autonomous Academic Support Systems, Explainable AI in Education, Knowledge Graph Reasoning, Multimodal AI Teaching Assistants