AgentX: Towards Agent-Driven Self-Iteration of Industrial Recommender Systems
AgentX employs a multi-agent system for autonomous self-iteration in industrial recommender systems, significantly boosting efficiency and business value.
Key Findings
Methodology
This study introduces AgentX, a multi-agent framework comprising four tightly integrated components: Brainstorm, Developing, Evaluation, and Harness Evolution, forming a closed-loop autonomous optimization cycle. The Brainstorm agent synthesizes evidence from historical experiments, system architecture, data analysis, and external research to generate a ranked set of executable proposals. The Developing agent converts these proposals into production-ready code using repository-grounded code generation and multi-dimensional reliability verification. The Evaluation agent safely rolls out changes in online A/B tests, judging outcomes with guardrail vetoes, and converts results into structured knowledge assets. The Harness Evolution layer employs semantic-gradient optimization (SGPO) to continuously update agent parameters based on accumulated trajectories, enabling self-improvement. This system leverages trajectory data to iteratively refine recommendation models and research strategies, demonstrating substantial improvements in iteration speed and business metrics.
Key Results
- In a three-week deployment on Kuaishou’s main feed and life-service recommendation scenarios, three AgentX workers autonomously generated 374 ideas, translating into 10 deployable solutions. The per-worker throughput doubled weekly through self-evolution, leading to an 8-fold increase in concurrency and a 3.7-fold business value uplift, with user app-time increasing by 0.561%. The annualized revenue exceeded RMB 100 million, validating the system’s effectiveness.
- The system achieved continuous self-evolution, with experiment throughput and online gains growing steadily over time, confirming the feasibility of closed-loop autonomous optimization at industrial scale.
- Beyond recommendation strategy improvements, AgentX supported autonomous reproduction of research papers, module ablation, and cross-paper architectural composition, demonstrating its versatility in both engineering and scientific domains.
Significance
AgentX marks a paradigm shift in industrial recommendation system development by transforming manual, linear iteration into an intelligent, self-evolving process. This approach drastically reduces reliance on human expertise, accelerates innovation cycles, and aligns system optimization directly with real business outcomes via online feedback. The closed-loop design ensures that each experiment’s trajectory data feeds back into the system, enabling continuous learning and adaptation. Such capabilities address long-standing bottlenecks in scaling recommendation R&D, paving the way for fully autonomous, scalable, and adaptive recommendation engines that can meet the demands of complex, multi-objective industrial environments.
Technical Contribution
The core technical innovation lies in integrating a multi-agent architecture with trajectory-based self-improvement mechanisms. The system’s four agents coordinate to generate hypotheses, implement code, evaluate results, and update models via semantic-gradient optimization. The use of repository-grounded code generation ensures high reliability, while SGPO enables continuous, data-driven parameter updates. This framework surpasses existing AutoML and agent-based approaches by enabling full lifecycle automation, real-time feedback integration, and long-term knowledge accumulation, thus establishing a new standard for autonomous recommendation system development.
Novelty
This work is the first to implement a fully autonomous, multi-agent, closed-loop recommendation system that leverages real online A/B feedback for continuous self-improvement. Unlike prior approaches limited to offline metrics or manual interventions, AgentX integrates real-time online signals and trajectory data to drive ongoing optimization. The introduction of semantic-gradient-based self-evolution represents a novel mechanism for agent parameter updates, enabling persistent learning. This comprehensive, end-to-end automation framework distinguishes itself from existing research by operationalizing autonomous R&D at industrial scale, demonstrating tangible business benefits.
Limitations
- While effective in large-scale environments, the system’s performance under extreme business shifts or data sparsity remains to be validated. Its adaptability to rapidly changing objectives or novel scenarios may require further tuning.
- The code generation and verification pipeline depend heavily on predefined repository structures, which could limit flexibility when facing radically new model architectures or highly customized deployment needs.
- The computational cost of continuous online experimentation and self-updating models is substantial, necessitating further optimization for resource efficiency, especially in resource-constrained settings.
Future Work
Future research will focus on enhancing the system’s robustness against extreme domain shifts and data scarcity, exploring more efficient code synthesis techniques, and integrating reinforcement learning to further improve autonomous decision-making. Additionally, expanding multi-objective optimization capabilities and reducing computational overhead will be key directions. The authors also plan to extend the framework to other industrial AI tasks such as personalized content generation and multi-modal recommendation, aiming to build a comprehensive autonomous AI development ecosystem.
AI Executive Summary
The landscape of industrial recommendation systems is characterized by complex, multi-faceted challenges. Traditional development processes rely heavily on manual effort, with engineers iterating through data analysis, feature engineering, model training, deployment, and evaluation in a cycle that often spans weeks. This manual approach is constrained by human cognitive limits and organizational bottlenecks, making it difficult to scale innovation and respond swiftly to dynamic business needs.
In response to these limitations, the authors propose AgentX, a pioneering multi-agent system designed to automate the entire lifecycle of recommendation model development and optimization. The system comprises four core agents—Brainstorm, Developing, Evaluation, and Harness Evolution—each responsible for a specific stage in the iterative process. These agents work collaboratively within a closed-loop framework, continuously generating hypotheses, translating them into production code, deploying them safely in online environments, and learning from the outcomes to inform subsequent iterations.
The Brainstorm agent acts as the system’s creative hub, synthesizing evidence from historical experiments, system architecture, data analysis, and external research to produce a ranked set of actionable proposals. It employs bounded exploration to manage ambiguity and ensure proposals are operationally feasible. The Developing agent then converts these proposals into high-quality production code, leveraging repository-grounded code generation techniques combined with multi-dimensional reliability checks to guarantee stability and correctness.
Once the code is deployed, the Evaluation agent manages safe online rollout, utilizing guardrail veto mechanisms to prevent adverse effects. It interprets A/B test results, transforming both successes and failures into structured knowledge assets that enrich the system’s understanding. The Harness Evolution layer employs semantic-gradient optimization (SGPO), a novel technique that iteratively refines agent parameters based on accumulated trajectories, fostering continuous self-improvement.
Empirical results from a three-week deployment on Kuaishou’s main feed and life-service recommendation scenarios demonstrate the system’s remarkable capabilities. The automated process generated 374 ideas, of which 10 were successfully launched, with per-worker throughput doubling weekly. The system achieved an 8-fold increase in concurrency, a 3.7-fold boost in business value, and a user app-time increase of 0.561%. These outcomes validate the potential of autonomous, self-evolving recommendation systems to transform industrial AI development.
Beyond immediate business gains, AgentX establishes a new paradigm for AI-driven R&D. By creating a fully closed-loop, trajectory-based optimization process, it enables continuous learning and adaptation, reducing reliance on manual intervention. The framework’s scalability and robustness suggest broad applicability across various industrial AI tasks, from content personalization to scientific research.
However, challenges remain. The system’s performance under extreme shifts, the computational costs associated with continuous online experimentation, and the flexibility of code generation pipelines warrant further investigation. Future work aims to enhance robustness, efficiency, and versatility, pushing the boundaries of autonomous AI in industry. Overall, AgentX exemplifies a significant step toward fully automated, self-improving recommendation engines capable of sustaining long-term business growth and technological innovation.
Deep Dive
Abstract
Recommendation algorithm iteration is moving from an artisanal, engineer-bound process toward an industrialized research loop, but this transition remains blocked by a structural execution bottleneck: the idea-to-launch cycle still depends on human engineers to generate hypotheses, modify production code, launch A/B experiments, and attribute online results. Innovation therefore scales linearly with headcount rather than compounding with evidence, compute, and accumulated experimental knowledge. We present AgentX, a production-deployed multi-agent system that fundamentally restructures this production function. AgentX operates as a self-evolving development engine: it autonomously generates, implements, evaluates, and learns from recommendation experiments at a scale and pace that no manual workflow can sustain. The system orchestrates four tightly coupled stages in a closed loop. A Brainstorm Agent synthesizes evidence from historical experiments, system architecture, data analysis, and external research into ranked, executable proposals. A Developing Agent translates each proposal into production-ready code through repository-grounded generation and multi-dimensional reliability verification. An Evaluation Agent conducts safe online rollout with guardrail-vetoed A/B judgment, converting both successes and failures into structured knowledge assets. A Harness Evolution layer (SGPO) then distills execution trajectories into semantic-gradient updates that continuously sharpen the agents themselves -- making the system not merely automated, but self-improving.
References (20)
RankMixer: Scaling Up Ranking Models in Industrial Recommenders
Jie Zhu, Zhifang Fan, Xiaoxie Zhu et al.
OneLive: Dynamically Unified Generative Framework for Live-Streaming Recommendation
Shenghui Wang, Yusheng Huang, Ruochen Yang et al.
KuaiRand: An Unbiased Sequential Recommendation Dataset with Randomly Exposed Videos
Chongming Gao, Shijun Li, Yuan Zhang et al.
Learning Tree-based Deep Model for Recommender Systems
Han Zhu, Pengye Zhang, Guozheng Li et al.
Auto-WEKA: combined selection and hyperparameter optimization of classification algorithms
Chris J. Thornton, F. Hutter, H. Hoos et al.
SORT: A Systematically Optimized Ranking Transformer for Industrial-scale Recommenders
Chunqi Wang, Bin Wu, Taotian Pang et al.
AutoML-Agent: A Multi-Agent LLM Framework for Full-Pipeline AutoML
Patara Trirat, Wonyong Jeong, S. Hwang
Towards Automated Neural Interaction Discovery for Click-Through Rate Prediction
Qingquan Song, Dehua Cheng, Hanning Zhou et al.
Gemma 2: Improving Open Language Models at a Practical Size
Gemma Team Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa et al.
R&D-Agent: An LLM-Agent Framework Towards Autonomous Data Science
Xu Yang, Xiao Yang, Shikai Fang et al.
GLM-5: from Vibe Coding to Agentic Engineering
GLM-4.5 Team Aohan Zeng, Xin Lv, Zhenyu Hou et al.
Self-EvolveRec: Self-Evolving Recommender Systems with LLM-based Directional Feedback
Sein Kim, S. Park, Hongseok Kang et al.
MixFormer: Co-Scaling Up Dense and Sequence in Industrial Recommenders
Xu Huang, Hao Zhang, Zhifang Fan et al.
Moment&Cross: Next-Generation Real-Time Cross-Domain CTR Prediction for Live-Streaming Recommendation at Kuaishou
Jiangxia Cao, Shenghui Wang, Yue Li et al.
Self-Evolving Recommendation System: End-To-End Autonomous Model Optimization With LLM Agents
Haochen Wang, Yi Wu, Daryl Chang et al.
Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations
Jiaqi Zhai, Lucy Liao, Xing Liu et al.
SMES: Towards Scalable Multi-Task Recommendation via Expert Sparsity
Yukun Zhang, Si-Thanh Dong, Xu Wang et al.
From Infrastructure to Culture: A/B Testing Challenges in Large Scale Social Networks
Ya Xu, Nanyu Chen, A. Fernandez et al.
DS-Agent: Automated Data Science by Empowering Large Language Models with Case-Based Reasoning
Siyuan Guo, Cheng Deng, Ying Wen et al.
AI-Researcher: Autonomous Scientific Innovation
Jiabin Tang, Lianghao Xia, Zhonghang Li et al.