AI Multi‑Agent System for E‑commerce: Design, Implementation, and Operational Insights

JD Retail has built a multi‑agent AI system, named Mario X, to serve as an intelligent merchant assistant that automates tasks ranging from product publishing to order management, customer service, and data analysis.

The system follows a Multi‑Agents approach based on the ReAct paradigm, where each LLM‑driven agent has a specific business role and can invoke tools or collaborate with other agents to solve merchant queries.

Real‑world merchant decision‑making is abstracted into three core roles: domain experts (mapped to individual agents), tools (atomic service APIs), and a "master" orchestrator (the "general manager"), which coordinates the workflow without deep domain expertise.

Agents operate through an inner reasoning loop consisting of a Thought step (natural‑language reasoning) and an Action Code step (executable instructions for tools or other agents). The Action Code includes fields such as dispatch target, input parameters, job description, and Trust_Mode to control further reasoning.

The hierarchical architecture separates a Master Agent that interacts directly with merchants from subordinate agents that execute specific tasks via APIs, enabling efficient scaling and rapid integration of new capabilities.

To ensure reliability, a full‑link ReAct monitoring framework collects Thoughts, Action Codes, and Observations, evaluates them with human labeling and large‑model scoring, and applies thresholds to halt or continue inference.

Training pipelines involve constructing vertical domain knowledge bases, generating SFT samples from real customer‑service data, enriching questions and actions, and pre‑training models to handle complex inputs, with reinforcement learning considered for open‑ended problem solving.

The paper concludes with a vision for a comprehensive AI‑driven merchant operating team that can flexibly combine various tools and expertise to improve experience, efficiency, decision quality, and cost for both merchants and the platform.