AgentX Automates Self-Iteration for Industrial Recommender Systems

The development and iteration of recommendation algorithms traditionally rely heavily on human engineers, creating a bottleneck where innovation scales linearly with staffing. This paper introduces AgentX, a multi-agent system designed to revolutionize this process by enabling autonomous, self-iterating recommender systems. AgentX is already deployed in production and aims to transform the idea-to-launch cycle into an industrialized, self-evolving development engine. AgentX operates through a closed loop involving four interconnected stages. A Brainstorm Agent generates and ranks executable proposals based on historical data and research. A Developing Agent then translates these proposals into production-ready code, complete with reliability verification. An Evaluation Agent conducts safe A/B experiments, converting outcomes into structured knowledge. Finally, a Harness Evolution layer (SGPO) continuously refines the agents themselves, ensuring the system not only automates but also self-improves, allowing innovation to compound rather than just scale.