AgenticRS: Agentic Recommender Systems

Large-scale recommender systems in industry commonly adopt a mature multi-stage architecture of recall, coarse ranking, fine ranking and re-ranking, and have evolved from collaborative filtering to deep neural networks and large pre-trained models. Nevertheless, both multi-stage pipelines and unified One Model designs remain essentially static: models are treated as black-box components, and system optimization chiefly relies on manually proposed hypotheses and iterative engineering, making it difficult to achieve continual, autonomous evolution under highly heterogeneous data and multi objective business constraints. This paper introduces the notion of an Agentic Recommender System (AgenticRS), which reorganizes decision units in recommendation through an agentic perspective. We distinguish between functional agents and model agents, and specify three constraints—functional closed loop, independent evaluability and evolvable decision space—to determine which modules should be designed as agents. For model agents, we present two complementary self evolution mechanisms: reinforcement learning based optimization when the action space is well-defined, and large language model (LLM) based generation and selection of new model and training schemes when the design space is complex and experience-driven. We further separate individual evolution and compositional evolution to describe both the improvement of a single agent and the evolving selection, connection and cooperation among multiple agents, and propose a layered Inner and Outer reward design to balance local capability optimization with global objective alignment. The proposed framework aims to provide a concise and general design foundation for transforming static recommendation pipelines into multi-agent, self-evolving recommender systems.