AgentX: Towards Agent-Driven Self-Iteration of Industrial Recommender Systems

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.

• The paper introduces AgentX, a self-iterating framework that integrates brainstorming, development, and evaluation agents to automate recommendation system improvements.
• It employs semantic-gradient prompt optimization to refine agent performance, achieving throughput enhancements and significant revenue gains in production deployment.
• The framework demonstrates practical viability through rigorous A/B testing and closed-loop feedback, shifting engineer roles from manual iteration to strategic oversight.

AgentX: An Agent-Driven Self-Evolving Framework for Industrial Recommender Systems

Motivation and Structural Innovation

AgentX addresses the bottlenecks inherent in industrial recommender system R&D, where iteration has historically relied on labor-intensive, engineer-driven processes. The transition from hypothesis generation to production rollout, A/B evaluation, and attribution remains blocked by sequential, manual handoffs that scale linearly with human effort. AgentX overcomes this through an agent-driven closed-loop architecture that autonomously generates, implements, evaluates, and iteratively refines recommendation strategies, leveraging trajectory data and real online feedback as compounding assets, rather than mere automation artifacts.

Figure 1: AgentX transforms recommendation iteration from a human-driven, manually handed-off pipeline into an agent-driven closed loop, where online A/B feedback and trajectory data continuously improve the system itself.

System Architecture

AgentX orchestrates four tightly integrated agents within a feedback-driven closed loop:

• Brainstorm Agent synthesizes evidence from historical experiments, system architecture, data analysis, and recent research to produce ranked, executable proposals, explicitly mitigating ambiguity via boundary-setting and evidence-weighted candidate scoring.
• Developing Agent grounds implementation in repository knowledge and toolkits, ensuring code generation is robust to feature and pipeline constraints, verified through deterministic checks, and scored across eight reliability dimensions, spanning semantic correctness to dryrun pipeline passes.
• Evaluation Agent chains safe production deployment, online A/B execution, and guardrail-vetoed judgment, treating real user feedback as the authoritative reward signal. Negative results are assetized to guide future search and pruning.
• Harness Evolution Layer (SGPO) applies semantic-gradient optimization to execution traces, updating agent prompts and harness specifications via paired replay for safe, inspectable subagent-level evolution.

  Figure 2: Overall framework of AgentX.

Agent Workflows and Verification Protocols

Brainstorm Agent

This agent not only proposes ideas but enforces operational precision through intake boundary normalization, batch proposal exploration classified by maturity (ready-to-implement, probe-first, moonshot-backlog), and evidence weighting across Experiment KB, System KB, Data Analysis, and Model Research. Candidate scoring integrates objective alignment, business validity, implementation feasibility, handoff completeness, and risk, with historical knowledge re-used to avoid known failure modes.

Figure 3: Brainstorm Agent workflow, showcasing evidence aggregation, proposal validation, and materialization leveraging Experiment KB and System KB.

Developing Agent

The Developing Agent realizes approved proposals into production-ready code, differentiated into online strategy and offline model tracks. Reliability is maintained through schema queries, DSL and syntax checkers, static linters, and repository knowledge. Implementation follows a staged plan-abstraction, atomic execution, verification, and dryrun loops, scored by weighted reliability metrics with severity factors prioritizing semantic correctness and default-off safety.

Figure 4: Architecture of the Developing Agent for the online strategy track, emphasizing repository-grounded generation and verification.

For offline model track, the Coding Agent enforces policy declaration, code verification, deterministic metric extraction, and attribution checking. Expert consensus and causal-chain validation are prerequisites for trustworthy conclusion assetization.

Figure 5: Architecture of the Coding Agent for the offline model track, combining policy declaration, code change verification, training execution, and expert attribution.

Evaluation Agent

This agent operationalizes safe deployment and traffic assignment, rigorous metric extraction, composite guardrail evaluation, and verdict assetization, integrating negative results as reusable exploration constraints.

Figure 6: Architecture of the Evaluation Agent, chaining OpenAPI-mediated deployment, A/B execution, and guardrail-vetoed judgment.

Harness Evolution via Semantic Gradient Prompt Optimization

AgentX's harness evolution utilizes SGPO, which distills trace-based evidence and rubric-encoded replay tasks into structured semantic gradients, updating subagent harness specifications and admitting edits solely through paired replay. Two modes—SGPO-I for brainstorm evolution, SGPO-II for coding harness refinement based on merge request replay—demonstrated improvements from 75% to 98% replay score and per-dimension gains from 2.60 to 4.90.

Figure 7: SGPO-I trace-based harness evolution, illustrating sampling, evaluation, refinement, and paired replay gate.

SGPO evolution is non-monotonic but robust by the admission gate, preventing regressive updates based on evaluator misdiagnosis.

Figure 8: Representative SGPO-I evolution patterns on brainstorm workflow—showing dynamic, non-monotonic improvement trajectories across subagents during harness evolution.

Figure 9: Case study of per-dimension score trajectory for a complex async module across SGPO-II self-evolution iterations.

Empirical Results: Throughput and Online Business Impact

Over a three-week production deployment on the Kuaishou App, AgentX demonstrated:

• 374 ideas processed by three agents, resulting in 10 launchable rollouts.
• Per-worker throughput doubled weekly; concurrent experiments quadrupled; business value per worker tripled compared to manual engineers.
• Cumulative app-time gain of 0.561% and over RMB 100M annualized revenue realized, with compounding returns from skill consolidation and template maturation.

  Figure 10: Weekly concurrent experiments quadrupled, idea pass rate tripled, and launchable results more than doubled through AgentX self-evolution.

Practical and Theoretical Implications

AgentX establishes that agent-driven, self-evolving industrial recommendation systems are not speculative but production-viable, delivering scalable and compoundable leverage. The division of labor shifts; engineers focus on evolving agent frameworks and selection, while agent systems drive the iterative loop autonomously. The trajectory data from every experiment fuels both engineering and agent system evolution, linking execution to long-term intelligence growth.

In offline model benchmark reproduction, AgentX consistently translated paper designs into executable modules, surfacing cross-dataset discrepancies and demonstrating genuine method transferability. Automated combinatorial module exploration further improved AUC on industrial datasets, validating the framework for both theoretical exploration and practical deployment.

Future Directions

Potential developments include:

• Extension of agentic closed loops to new business verticals and non-recommendation ML domains.
• Integration of more granular traffic and business feedback into agent evolution.
• Advancement of foundation models via continuous agent-data-driven pretraining.
• Broader adoption of agent-driven automation across R&D infrastructure, reducing structural bottlenecks and increasing leverage beyond human headcount.

Conclusion

AgentX demonstrates a robust, agent-driven closed-loop architecture for industrial recommender system evolution, rebalancing the production function from linear human labor to compoundable agent leverage. With strong empirical gains, rigorous trace-based harness evolution, and practical deployment at scale, AgentX invites further exploration and adoption of agentic automation in recommendation and broader AI engineering domains.

Figure 11: Production flow for the life-service recommendation decision agent—AgentX converts diagnostic signals into production operators, expanding atomic action space available to the decision agent.

(2606.26859)