A production-grade multi-agent assistant built on architecture, designed for cross-border e-commerce scenarios.
개요
A production-grade multi-agent assistant built on architecture, designed for cross-border e-commerce scenarios. While e-commerce serves as the reference domain, the architecture is fully extensible to other conversational AI use cases. This system adopts a heterogeneous architecture, deeply integrated with the Model Context Protocol (MCP) specification and built on the latest LangChain 1.x design principles. It provides a with highly modular and extensible middleware lifecycle hooks, suitable for complex agent applications and task orchestration scenarios. The system implements three middlewares on top of LangChain Deep Agent's middleware mechanism, covering , , and . They are mounted on the Orchestrator in sequence and intervene at different lifecycle hooks. * A custom summarization strategy is used instead of the built-in SummarizationMiddleware to allow fine-grained control over summarization timing and storage logic.
README
Trade Agent Skills — Cross-Border E-Commerce Multi-Agent Assistant
A production-grade multi-agent assistant built on LangChain 1.2.6 + Deep Agents + AgentSkills architecture, designed for cross-border e-commerce scenarios. While e-commerce serves as the reference domain, the architecture is fully extensible to other conversational AI use cases.
Tech Stack (Java + Python)
This system adopts a Java (MCP Server) + Python (Agent Brain) heterogeneous architecture, deeply integrated with the Model Context Protocol (MCP) specification and built on the latest LangChain 1.x design principles.
It provides a production-grade Agent Skills reference architecture with highly modular and extensible middleware lifecycle hooks, suitable for complex agent applications and task orchestration scenarios.
| Layer | Technology | Description |
|---|---|---|
| Agent Framework | LangChain 1.2.6 + Deep Agents 0.3 + LangGraph 1.0.7 | Core orchestration & multi-agent collaboration |
| LLM Engine | Alibaba Cloud Bailian (customizable) | Swap in any LLM provider as needed |
| Backend (Python) | FastAPI + Uvicorn + SQLAlchemy | Agent Brain API & services |
| Backend (Java) | Spring Boot 3.4.1 + Spring AI 1.0.0 | MCP Server tool services |
| Database | MySQL 8.0+ | Business data persistence |
| Cache / State | Redis | Session management & ephemeral state |
| Vector Search | Milvus (BM25 + Dense) | Hybrid retrieval for memory augmentation |
| ORM (Java) | MyBatis-Plus 3.5.9 | Java-side data access |
| Tool Protocol | MCP (Model Context Protocol) | Cross-language tool invocation standard (SSE) |
| Email Service | Spring Mail (SMTP) | Automated email notifications |
| Evaluation | agentevals + LLM-as-Judge | Automated trajectory & quality assessment |
Architecture Overview
┌───────────────────────────────────────────────┐
│ FastAPI Gateway │
└────────────────────┬──────────────────────────┘
│
┌───────────▼────────────┐
│ Deep Agent Orchestrator│
│ Planning / Skills / │
│ FileSystem Backend │
└───────┬────────────────┘
│ On-demand dispatch
┌────────────┼────────────────────────┐
▼ ▼ ▼ ▼
┌────────┐ ┌─────────┐ ┌─────────┐ ┌──────────┐
│ Order │ │Logistics│ │Comms │ │Analytics │
│ Agent │ │ Agent │ │ Agent │ │ Agent │
└───┬────┘ └───┬─────┘ └───┬─────┘ └───┬──────┘
└──────────┴─────┬─────┴────────────┘
│
┌────────▼────────┐
│ MCP Protocol │
│ (SSE Client) │
└────────┬────────┘
│
┌───────────▼────────────┐
│ Java MCP Server │
└────────────────────────┘
Middleware Layer Design
The system implements three middlewares on top of LangChain Deep Agent’s middleware mechanism, covering memory management, message persistence, and response quality guardrails. They are mounted on the Orchestrator in sequence and intervene at different lifecycle hooks.
Request Incoming
│
▼
MemoryMiddleware.before_agent ← Restore history + retrieve context
PersistenceMiddleware.before_agent ← Persist user message
│
▼
[Agent Reasoning / Tool Calls]
│
▼ (on each LLM call)
MemoryMiddleware.before_model ← Trigger summarization if needed *
QualityGuardMiddleware.wrap_model ← Evaluate & retry low-quality responses
│
▼
PersistenceMiddleware.after_agent ← Persist AI response
* A custom summarization strategy is used instead of the built-in
SummarizationMiddlewareto allow fine-grained control over summarization timing and storage logic.
Human-in-the-Loop: Email Confirmation
Email sending requires explicit human approval, implemented via LangGraph’s interrupt():
User requests email → Agent drafts → interrupt() pauses → Frontend shows preview
↓
User: approve / reject / edit
↓
POST /resume → Command(resume=decision) → Resume execution
Key Features
| Feature | Description |
|---|---|
| AgentSkills | 7 domain skills loaded on demand to reduce token overhead |
| SubAgent Delegation | Complex tasks are automatically dispatched to specialized sub-agents |
| Planning Tool | Built-in task planner that decomposes multi-step requests |
| FileSystem Context | Virtual file system for managing long documents and analysis reports |
| Hybrid Memory | MySQL persistence + Milvus vector retrieval + Redis session cache |
| MCP Tool Calls | Connects to the Java backend via SSE protocol |
| Human-in-the-Loop | Mandatory human confirmation before sending emails |
| 4-Layer Evaluation | Trajectory + Quality + Safety + LLM-as-Judge |
Project Structure
trade-agent-brain/
├── app/
│ ├── agents/
│ │ ├── orchestrator.py # Deep Agent orchestrator
│ │ └── subagents.py # Sub-agent definitions
│ ├── config/
│ │ ├── settings.py # Configuration management
│ │ ├── llm_config.py # LLM configuration
│ │ ├── database.py # MySQL connection
│ │ └── redis_config.py # Redis connection
│ ├── middleware/
│ │ ├── memory_middleware.py # Unified memory middleware
│ │ ├── persistence_middleware.py # Persistence middleware
│ │ └── quality_guard_middleware.py # Quality guard middleware
│ ├── models/ # Data models
│ ├── routers/ # API routes
│ ├── services/ # Business services
│ ├── tools/ # MCP tools
│ └── main.py # FastAPI entry point
├── skills/ # AgentSkills directory
│ ├── order-management/
│ │ └── SKILL.md
│ ├── logistics-tracking/
│ │ └── SKILL.md
│ ├── cart-management/
│ │ └── SKILL.md
│ ├── email-notification/
│ │ └── SKILL.md
│ ├── customs-clearance/
│ │ └── SKILL.md
│ ├── data-analytics/
│ │ └── SKILL.md
│ └── dispute-resolution/
│ └── SKILL.md
├── tests/
├── requirements.txt
├── .env.example
└── README.md
trade-mcp-server/ # Java MCP Server (Maven multi-module)
├── pom.xml
├── mcp-common/
│ ├── pom.xml
│ └── src/main/java/com/cbec/mcp/common/
│ ├── entity/
│ ├── enums/
│ ├── result/
│ └── util/
├── mcp-server/ # MCP Server main module
│ ├── pom.xml
│ └── src/main/
│ ├── java/com/cbec/mcp/server/
│ │ ├── McpServerApplication.java
│ │ ├── config/
│ │ │ └── McpConfig.java # Unified MCP tool registration
│ │ ├── dto/ # Data transfer objects
│ │ ├── mapper/ # MyBatis-Plus mapper interfaces
│ │ ├── service/ # Domain service layer
│ │ └── tool/ # MCP Tool definitions (@Tool)
│ └── resources/
│ ├── application.yml
│ └── mapper/
└── sql/ # Database scripts
sql/ # Global SQL scripts
├── schema.sql # Table creation script
└── data-demo.sql # Demo data
trade-mcp-server
Module Overview
- mcp-common — Shared layer containing database entities, enums, the unified response wrapper
McpResult, and JSON utilities. Depended on bymcp-server. - mcp-server — Core service module containing MCP Tool definitions, service logic, MyBatis mappers, and the Spring Boot entry point.
MCP Tools
All tools are registered centrally via McpConfig using the @Tool annotation, and are automatically exposed on the SSE endpoint for MCP Client discovery and invocation.
Connecting to trade-agent-brain
The Python side (trade-agent-brain) connects to trade-mcp-server via MCP SSE Client. The connection URL is configured in .env:
MCP_SERVER_URL=http://127.0.0.1:8081/sse
MCP_CALL_TIMEOUT=30
app/tools/__init__.py defines a generic call_mcp_tool() function that establishes an SSE connection using mcp.client.sse.sse_client and invokes remote Java-side @Tool methods via ClientSession.call_tool(). The full call chain is:
Agent Reasoning → Python @tool → MCP SSE Client → Java MCP Server (SSE endpoint)
→ @Tool method → Service → MyBatis Mapper → MySQL → McpResult JSON response
Getting Started
🚧 The project is still under active development. A full setup guide will be provided soon.
License
추천 도구
다른 키워드를 입력하거나 필터를 제거해 보세요.
설치
npx skillfish add hungrycodingsir/trade-agent-skills