Internal E-commerce Operations Copilot: A Practical Guide to Multi-Agent Collaborative Workflow Based on n8n

This is an open-source project demonstrating the implementation capabilities of enterprise-level AI workflows, maintained by talon1126. The source code is available on GitHub (https://github.com/talon1126/agent). The project orchestrates multi-departmental agent workflows via n8n, builds AI services using FastAPI, simulates e-commerce operation scenarios such as warehousing, procurement, logistics, and customer service, embodies the engineering practice of AI Ops, and provides a reference for the implementation of enterprise-level AI applications.

With the improvement of Large Language Model (LLM) capabilities, enterprises are exploring the integration of AI agents into business processes. However, there is an engineering gap from Proof of Concept (POC) to production environment: how to orchestrate multi-agent collaboration, integrate with existing systems, and ensure decision-making is controllable and auditable? This project, set against the backdrop of a fictional internal e-commerce platform, uses n8n, FastAPI, and Feishu to build a multi-department collaborative AI agent system to address these issues.

The project adopts a layered architecture: User Message → Feishu Gateway Adapter → n8n Workflow → AI Service → Mock Enterprise API → Postgres Database. Core components include:

1. Feishu Gateway Adapter: A multi-department robot gateway that supports message routing and deduplication;
2. n8n Workflow Engine: The central hub for agent orchestration, with independent workflows for each department;
3. AI Service (FastAPI): The intelligent decision-making layer, providing endpoints like /decide and schema validation;
4. Mock Enterprise API: Simulates business systems such as orders and inventory, supporting logging and replay.

Functions of each department's agent:

• Warehousing Agent: Manages inventory by batch + location, supports query, anomaly detection, and replenishment application;
• Procurement Agent: Handles replenishment approval, purchase order generation, and arrival confirmation;
• Logistics Agent: Handles delivery status query and anomaly processing;
• Customer Service Agent: Handles order query, refund processing, and work order creation.

AI Ops practices embodied in the project:

1. Schema Validation and Type Safety: AI outputs undergo strict schema validation, and tool parameters are verified using Pydantic;
2. Approval Guardrails: High-risk operations (high-value refunds, large-scale procurement) require manual approval;
3. Operation Logs and Observability: Structured logs record the entire execution process;
4. Dead Letter Queue and Replay Mechanism: Failed events enter the dead letter queue, supporting replay recovery.

Three typical scenarios:

1. Inventory Alert and Automatic Replenishment: Warehouse scans inventory → Creates replenishment application → Procurement approval → Generates purchase order → Synchronizes arrival;
2. High-Value Refund Approval: Customer applies → Customer service judges threshold → Creates approval → Manual approval → Executes refund;
3. Automatic Follow-up for Logistics Anomalies: Logistics monitoring detects anomaly → Creates follow-up case → Notifies customer service → Synchronizes status via Feishu.

Value Insights: Progressive implementation, human-machine collaboration, observability, scalability, localization adaptation; Limitations: In-memory storage (needs Postgres persistence), mock AI service (can connect to real LLM), lack of permission management and RAG; Summary: This project is a complete enterprise-level AI workflow case, demonstrating the construction of a multi-agent system using open-source toolchains. The AI Ops practices have reference value for enterprise AI implementation.