Agentic AI for Recommender Systems: A Multi-Agent Framework for Grounded, Explainable, and Constraint-Aware Movie Recommendation

Abstract

Large language models (LLMs) are increasingly used as reasoning engines within multi-step, tool-using agent pipelines, yet their value for core recommendation tasks relative to classical collaborative filtering (CF) remains poorly characterized. We present an agentic recommender system for the MovieLens-100K dataset that decomposes recommendation into four cooperating agents: a Planner that converts free-text user requests into structured intent, a Retriever that performs deterministic, tool-based candidate generation via singular value decomposition (SVD), a Ranker/Explainer that selects and justifies a final list using a locally hosted Qwen2.5-3B-Instruct model, and a Critic that verifies genre-exclusion constraints and can trigger a bounded retry loop. We evaluate the system along four axes: ranking accuracy (Hit@K, NDCG@K) against Random, Popularity, and CF-only baselines; constraint-satisfaction rate under free-text exclusion requests; automatic groundedness of generated explanations; and end-to-end latency. On a leave-one-out protocol with n = 600 evaluation users, the agentic pipeline achieves Hit@10 = 0.120 and NDCG@10 = 0.059, closely approaching the pure SVD baseline (Hit@10 = 0.130, NDCG@10 = 0.060) while substantially outperforming Popularity (Hit@10 = 0.030) and Random (Hit@10 = 0.002) baselines. Uniquely, the agentic pipeline achieves a 0.28% constraint-violation rate across 712 recommendations spanning 500 constrained queries, and 95.05% of generated explanations are automatically verified as grounded in the recommended movie's true genre metadata – capabilities that CF-only, Popularity, and Random baselines cannot provide by construction. We report this latency overhead honestly (≈ 23.7s per user versus ≈ 0.001s for CF-only, a ∼19,300× difference) and discuss the implications for deployment. Our findings suggest that LLM-based re-ranking of an already-strong retriever does not straightforwardly improve ranking accuracy, but adds genuinely new, verifiable capabilities – constraint adherence and natural-language explanation – that are structurally unavailable to classical recommenders.