From scoring to workflow: A survey of agentic recommender systems

Agentic recommender systems (ARS) extend recommendation from one-shot item scoring to interactive, adaptive, and feedback-driven workflows. Yet existing surveys rarely examine the coupling between agentic mechanisms and recommender system pipelines. This survey provides a recommender-system-centered analysis of ARS by examining how agents restructure decision objects across the workflow. Through systematic literature search, this survey constructs an analytical corpus of 47 studies published or made publicly available between January 2023 and April 2026. It organizes current ARS research into four workflow stages: User Modeling, Adaptive Decision Orchestration, Candidate Optimization, and Exposure Refinement. Across these stages, this survey synthesizes representative methods, evaluation resources, design trade-offs, deployment burdens, cross-stage dependencies, and recurring failure modes. The analysis also identifies key challenges in simulator fidelity, memory scalability, strategy-candidate alignment, candidate provenance, explanation faithfulness, feedback reliability, and runtime governance. It also highlights future opportunities enabled by frontier agentic mechanisms. Overall, this survey offers an RS-centric framework for understanding current progress and guiding reliable, scalable, and governable ARS research.