How Multi‑Agent AI Is Revolutionizing E‑Commerce Decision Making

Real‑world Decision Process

Merchants receive platform information such as new rules or penalties and need a consulting advisor to interpret these messages, extract key issues, and guide problem‑solving through knowledge bases, search, and human summarization.

Mapping Real Advisors to Agents

Instead of creating separate agents for each staff role, JD builds a single AI Agent that represents the collective expertise of online customer service, operations staff, and product managers, providing an always‑online encyclopedia for merchants while reducing platform costs.

AI‑Powered Merchant Team

The merchant team consists of domain experts, tools, and a "general manager" agent that orchestrates resources. The general manager receives merchant queries, dispatches tasks to specialized agents (e.g., data analysts), aggregates their outputs, and returns concise recommendations.

ReAct Agent Construction

Each agent follows an inner loop of reasoning (Thought) and Action Code generation. Thought describes the solution strategy in natural language; Action Code translates this into executable commands for tools or other agents, specifying the target, inputs, job description, and trust mode.

Multi‑Agent Workflow

When a merchant asks about deposit requirements, the Master Agent retrieves relevant memory, asks clarifying questions via an Echo tool, then delegates to a Consulting Advisor agent that queries the appropriate API. Subsequent rounds may involve name‑generation APIs or other specialized tools, with Trust_Mode controlling whether further reasoning is needed.

Layered Architecture

The architecture decomposes large‑model generation into hierarchical steps, reducing individual model complexity and enabling rapid iteration and easy integration of new agents such as marketing assistants.

Key Deployment Technologies

JD has launched a multi‑agent assistant for merchant onboarding, handling deterministic queries (e.g., platform rules) with high accuracy, while planning to incorporate reinforcement learning for more open‑ended business problems.

ReAct SFT: Domain Sample Construction

Domain knowledge is collected from real customer‑service logs, cleaned, and organized into a path tree. Scripts describing possible Q&A scenarios are created, then expanded with thought and action annotations to form a supervised fine‑tuning (SFT) dataset.

Complex Input Thought Generation

For ambiguous queries, the agent first generates a Thought, then refines the Action Code to query the correct knowledge base (e.g., mapping "red wine" to the "wine" category) before producing a final answer.

Full‑Chain ReAct Monitoring

Thoughts, Action Codes, and Observations are logged, manually labeled, and evaluated with large models to compute a ReAct score. Low‑scoring agents are halted, while successful agents continue through subsequent reasoning rounds.

Future Outlook

JD aims to extract human business reasoning into large models, integrate reinforcement learning, and build a flexible toolchain that lets merchants combine supply‑chain, pricing, and selection utilities according to their own workflow.