AgenticRS-Architecture: System Design for Agentic Recommender Systems

A research team led by Hao Zhang and Jinxin Hu has published a paper proposing 'AutoModel', a novel agentic architecture designed to overhaul the lifecycle of industrial-scale recommender systems. Instead of relying on traditional, fixed recall-and-rank pipelines, AutoModel organizes the recommendation process as a set of interacting, autonomous agents. These agents possess long-term memory and self-improvement capabilities, enabling the system to evolve continuously. The architecture is built around three core agents: AutoTrain for automated model design and training, AutoFeature for intelligent data analysis and feature evolution, and AutoPerf for managing performance, deployment, and online A/B testing. A central coordination and knowledge layer connects these agents, logging all decisions and outcomes to create a persistent memory for the system.

In a practical demonstration, the researchers detailed a module within AutoTrain called 'paper autotrain'. This module automates the complex process of translating a new machine learning method from an academic paper into a trained, production-ready model. It closes the loop from parsing the paper's methodology, to generating executable code, conducting large-scale training, and performing offline evaluations. This drastically reduces the manual engineering effort required for method transfer, accelerating innovation. The authors argue that AutoModel enables a paradigm of 'locally automated yet globally aligned' evolution for massive systems like those at Meta or TikTok, and its principles could generalize to other AI domains like search and advertising.