AI Multi‑Agent System for E‑commerce Merchant Assistance: Design, ReAct Architecture, and Implementation

JD Retail has built an AI‑powered e‑commerce assistant called the "Merchant Smart Assistant" that combines multiple decision‑making functions—from product publishing to order management, customer service, and data analysis—into a single tool.

The core of the system follows a Multi‑Agents approach based on the ReAct paradigm, where several specialized LLM agents (each with a dedicated business role) cooperate through tool calls and inner‑loop reasoning to solve merchant queries.

In practice, merchants receive a consulting‑style knowledge base from the platform; the assistant abstracts this into a single AI agent that represents the collective expertise of online客服, operations staff, and product managers, providing 24/7, cost‑effective support.

The assistant’s team structure is broken down into three role types: Domain Experts (agents that embody decision‑making and tool‑dispatch capabilities), Tools (atomic service APIs without decision power), and a Manager (a generalist agent that orchestrates the workflow and possesses broad e‑commerce knowledge).

JD Retail’s AI merchant team, named Mario X , is led by a Master Agent that coordinates multiple domain agents and accesses a suite of tool APIs. Benefits include continuous availability, higher efficiency, improved decision quality, and reduced operational costs.

Each ReAct agent performs an inner loop consisting of Thought generation (natural‑language reasoning) followed by Action Code creation (machine‑readable commands). The Action Code includes fields such as dispatch object, input parameters, job description, and Trust_Mode (which determines whether further reasoning is required).

Example workflow: a merchant asks about the required deposit for opening a JD store. The Master Agent retrieves relevant user features, realizes additional context is needed, and calls an Echo tool to ask the merchant for the product category. After the merchant replies “flowers”, the Master Agent creates a job description for a Consulting Advisor agent, which queries the appropriate deposit via a dedicated API. In a third round, the Master Agent uses a "Shop Name Generator" API to propose store names, demonstrating a three‑step ReAct process with varying Trust_Mode values.

The architecture is hierarchical: a large model’s complex generation task is decomposed into multiple, smaller reasoning steps, reducing model size and enabling rapid iteration and easy integration of new agents (e.g., marketing agents).

Potential challenges include error propagation from the Master Agent and longer latency due to multi‑step reasoning. JD Retail mitigates these with full‑link monitoring that captures Thought, Action Code, and Observation, then evaluates them using human labeling and large‑model scoring.

Key implementation milestones include launching a dedicated admission‑assistant agent, handling deterministic queries (e.g., platform rule clarifications), constructing vertical domain SFT samples from real customer‑service logs, training models to generate high‑quality Thought for complex inputs, and establishing a monitoring pipeline that halts low‑scoring agents early.

Looking forward, JD Retail aims to extract human business reasoning into large models, incorporate reinforcement learning, and build a flexible tool‑orchestration framework that lets merchants combine supply‑chain, selection, pricing, and other capabilities according to their own workflows.

Recommended reading links are provided at the end of the original article.