AgentX recommender: Agent-driven self-iteration system

AgentX is a production-deployed multi-agent system that autonomously generates, implements, evaluates, and learns from recommendation experiments. The paper was submitted to arXiv on 25 Jun 2026 as arXiv:2606.26859 and lists Changxin Lao and 59 other authors; the submission file is recorded as 6,578 KB.

How does AgentX work?

AgentX runs a closed loop of four tightly coupled stages: a Brainstorm Agent, a Developing Agent, an Evaluation Agent, and a Harness Evolution layer (SGPO). The Brainstorm Agent synthesizes evidence from historical experiments, system architecture, data analysis, and external research into ranked, executable proposals. The Developing Agent translates proposals into production-ready code using repository-grounded generation and multi-dimensional reliability verification. The Evaluation Agent performs safe online rollout with guardrail-vetoed A/B judgment and converts successes and failures into structured knowledge assets. The Harness Evolution layer distills execution trajectories into semantic-gradient updates that continuously sharpen the agents.

Each stage has an operational role: idea generation, code production, safe online testing, and agent self-improvement. The paper frames AgentX as what it calls a "self-evolving development engine," and emphasizes that the loop converts experimental outcomes into reusable knowledge assets. The system targets the core bottleneck the authors identify: the idea-to-launch cycle still depends on human engineers to generate hypotheses, modify production code, launch A/B experiments, and attribute online results.

What components and outputs does the system produce?

AgentX organizes work around agents plus artifact outputs: brainstormed proposals, production-ready code, guardrailed A/B results, and structured knowledge assets for future iteration. The Brainstorm Agent produces ranked, executable proposals. The Developing Agent produces repository-grounded code coupled with reliability checks. The Evaluation Agent issues guardrail-vetoed A/B judgments and records both successes and failures. The Harness Evolution layer (SGPO) produces semantic-gradient updates that feed back into the agents, closing the loop.

The authors position the Evaluation Agent as responsible for safe online rollout and for converting outcomes into structured knowledge assets. That conversion is explicit in the abstract: the Evaluation Agent converts both successes and failures into structured knowledge assets. The Harness Evolution layer, labeled SGPO in the paper, then distills execution trajectories into semantic-gradient updates.

Why does this matter?

AgentX addresses a scaling problem the authors describe: iteration on recommender systems currently scales linearly with headcount because human engineers must drive the entire idea-to-launch cycle. By automating hypothesis synthesis, code generation, guarded rollout, and self-improvement, AgentX aims to change that production function so iteration can compound with evidence, compute, and accumulated experimental knowledge rather than just headcount. That shift matters for teams that run large numbers of A/B experiments and want to capture failure as well as success in reusable assets.

What to watch

Follow whether the system moves beyond the lab description to broader industry deployments and whether the SGPO mechanism demonstrably improves agent proposals over successive cycles. The arXiv submission identifies the system concept and pipeline; the next concrete signals to look for are reproducible case studies or public release of code or evaluation traces from production rollouts.

Additional details The paper is categorized under Artificial Intelligence (cs.AI), Computation and Language (cs.CL), and Information Retrieval (cs.IR). Authors are listed alphabetically by their first name, as noted in the submission comments. The document is available on arXiv as arXiv:2606.26859, submitted 25 Jun 2026.