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

🕒 Published (v1): 2026-06-25 10:42 UTC · Source: Arxiv · link

Why this paper was selected

Industrial closed-loop agent self-iterating recommender systems; idea-to-launch automation

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

AgentX is a production-deployed multi-agent system at Kuaishou that closes the full idea-to-A/B-result loop for industrial recommender system iteration autonomously. It orchestrates four coupled agents (Brainstorm → Develop → Evaluate → Harness Evolution) and uses a Semantic-Gradient-based Prompt Optimization (SGPO) layer to self-improve from execution trajectories. In a three-week live deployment, three AgentX workers converted 374 ideas into 10 launchable rollouts with throughput doubling weekly.

Problem¶

Industrial recommender iteration is structurally bottlenecked by human engineers who must manually perform data analysis, code modification, A/B deployment, and result attribution—making throughput scale linearly with headcount. Existing LLM-AutoML and agentic ML systems fail three requirements for production: they rely on offline metrics rather than real online A/B feedback, are validated only at academic scale (not billion-sample, multi-objective production systems), and lack any continuous self-improvement mechanism that compounds over repeated deployments.

Method¶

AgentX runs a four-stage closed loop:

Brainstorm Agent: Converts a vague optimization intent into ranked, executable proposals via bounded batch exploration. Candidates are classified into three maturity states (ready-to-implement, probe-first, moonshot-backlog). Evidence is drawn from four sources—Experiment KB (historical launch reviews), System KB (architecture/feature/code scope), Data Analysis (SQL-based metric checks), and Model Research (paper-derived mechanisms)—with candidate-specific evidence weighting: for candidate \(c\) and structured boundary \(q\), the evidence set \(\mathcal{K} = \{\text{Experiment KB, System KB, Data Analysis, Model Research}\}\) is accessed through a weighting mechanism matched to the candidate's type.
Developing Agent: Translates proposals into production-ready code via repository-grounded generation (code context retrieved from the actual codebase) and a verification-oriented multi-dimensional implementation loop including quality scoring and failure-mode cataloging. Also handles model-side development: paper reproduction, module ablation, and cross-paper architectural composition.
Evaluation Agent: Manages safe online rollout with progressive traffic assignment, applies guardrail vetoes to A/B results (blocking launches that violate business safety thresholds), and converts both positive and negative outcomes into structured knowledge assets ("negative-result assetization").
Harness Evolution (SGPO): Reasons over accumulated execution trajectories to produce semantic-gradient updates to agent prompts. Two SGPO variants (SGPO-I and SGPO-II) update subagent specifications; updates are admitted only after paired replay validation to prevent regression.

A shared Data Layer persists all KB entries and experiment reports; a Monitoring Platform tracks system health.

Key Contributions¶

End-to-end closed loop for industrial recommendation: idea → code → A/B → trajectory feedback, deployed in production at Kuaishou without human stitching between stages.
Real online A/B signal as the reward: no reliance on offline metrics or human ratings; business value (user app-time, revenue) is the optimization criterion.
Self-evolving flywheel via SGPO: execution trajectories continuously refine agent prompts and foundation models, producing compounding (not linear) throughput gains.
Maturity-stratified proposal generation (ready/probe/moonshot) with residual avoid-set contraction across brainstorm rounds.
Negative-result assetization: failed experiments are structured into reusable failure memory rather than discarded.
Extension to model-side research: autonomous paper reproduction, module ablation, and cross-paper architectural composition within the same closed-loop principle.

Results¶

374 ideas → 10 launchable rollouts across three weeks with three AgentX workers on Kuaishou main-feed and life-service recommendation.
Per-worker throughput doubled weekly via self-evolution.
8× concurrency over a single manual engineer.
3.7× business value over a manual engineer.
+0.561% user app-time gain measured online.
>RMB 100M annualized revenue attributed to AgentX-launched strategies.

Limitations¶

Evaluation is a single three-week deployment at one company (Kuaishou); generalizability to other recommendation platforms or domains is undemonstrated.
The reward signal (online A/B) is delayed and non-stationary, creating inherent lag in the self-improvement loop; the paper acknowledges this as a "real-world reward problem" but does not fully resolve it.
SGPO prompt updates are admitted only through paired replay, which introduces a validation bottleneck that may slow adaptation when trajectory volume is low.
The system still requires human review in the Brainstorm Agent validation step and human oversight for guardrail configuration—it is not fully autonomous end-to-end.
Industrial-scale results are not compared against other agentic ML systems under controlled conditions; baselines are manual engineers only.

Relevance to Agentic AI / LLM Agents¶

AgentX is a rare example of a multi-agent system that completes the full outer loop of an engineering discipline—not just automating individual steps but closing the feedback cycle from intent to real-world reward and back to agent self-improvement. The SGPO mechanism is directly relevant to work on automatic prompt optimization and agent self-refinement, but grounds it in production business signals rather than synthetic benchmarks. The maturity-stratified brainstorming, negative-result assetization, and guardrail-vetoed evaluation are engineering patterns transferable to other agentic pipelines operating in high-stakes production environments. For researchers tracking LLM agents, this paper provides strong evidence that closed-loop, self-evolving agents can deliver compounding returns at industrial scale—addressing a standing critique that agentic systems have only been validated in sandbox settings.