Production-Grade AI Agent Architecture Practice: Market Intelligence System Based on LangGraph

This project presents a complete production-ready AI agent architecture. It achieves multi-step reasoning, tool calling, and human-machine collaboration via LangGraph's cyclic state graph, integrating technologies like multi-source RAG and MCP protocol to solve the challenges of building autonomous yet controllable AI systems, providing a practical and implementable model for real-world scenarios such as financial market intelligence.

AI agents are moving from proof-of-concept to production deployment, but balancing autonomous decision-making with safety and controllability remains a challenge for developers. This project targets real-world financial research scenarios, supporting automatic collection of market data, financial report analysis, internal document retrieval, and requesting human confirmation before sensitive operations. The tech stack includes Python3.12+, LangGraph (state management), FastAPI (asynchronous API), Streamlit (interactive interface), Docker (containerization), and can be seamlessly migrated to cloud environments like AWS App Runner.

1. Cyclic State Graph: Replace traditional linear pipelines, support multi-step reasoning, self-correction, and tool calling. Use AsyncSqliteSaver for state persistence to ensure breakpoint recovery after service restart;
2. MCP Protocol: Unify access interfaces for heterogeneous tools, integrate MCP servers like SQLite (CRM data), yfinance (stock prices), Filesystem (files), Playwright (browser) to enhance scalability and replaceability;
3. Dual-Layer Memory System: Short-term memory (BaseStore saves session facts) and long-term memory (InMemoryStore persists across sessions, planned upgrade to AsyncSqliteStore), managed via MCP tools.

1. Human-in-the-Loop (HITL): Classify tools into read-only (auto-executed) and side-effect (user approval required) categories, implement sensitive operation confirmation via LangGraph's dynamic interruption mechanism;
2. Multi-Source RAG: Static knowledge base (Pinecone vector database + OpenAI embedding model) stores internal documents; dynamic information is obtained via yfinance, Tavily search, Playwright;
3. Headless Browser: Playwright supports accessing dynamic web pages, obtaining accessibility trees, and saving screenshots for evidence;
4. Sandbox Workspace: The data/workspace directory provides file I/O capabilities, with write operations controlled by HITL.

Demo Scenario: Query VIP customer Yaniv Bohbot in CRM → Retrieve Amazon's 2024 AI strategy → Send email (user confirmation required), covering structured query, vector retrieval, content generation, human-machine interaction, etc. Deployment Operations: Docker containerization ensures environment consistency; uv accelerates dependency installation; environment variables manage sensitive configurations; FastAPI and Streamlit are separated to support independent scaling.

This project is an implementable model for production-grade AI agents, organically combining LangGraph state management, MCP standardization, RAG knowledge retrieval, and HITL security control. Key insights: Architecture design is more important than model parameters; HITL is essential for production deployment in high-risk scenarios; standardized protocols reduce tool management complexity; state persistence ensures system reliability. It has important reference value for developers exploring agent productionization.