Building a Production‑Ready Multi‑Agent System with LangGraph, MCP, A2A, and Ollama
Build a multi‑agent system using LangGraph, MCP, A2A, and Ollama, and capture traces with Langfuse while evaluating quality with DeepEval.
Set up LangGraph, MCP, A2A, and Ollama, then run the provided code to observe agent orchestration and trace logging with Langfuse.
Summary
The handbook presents a production‑ready multi‑agent system built with LangGraph, MCP, A2A, and Ollama, demonstrating how to orchestrate agents, standardize tool access, coordinate across frameworks, and capture observability. It starts by explaining the four‑agent architecture: a Curriculum Planner, Human Approval, Explainer, Quiz Generator, and Progress Coach, all connected via LangGraph workflows. MCP servers provide a protocol for agents to call external tools, while A2A services enable cross‑framework delegation, such as calling a CrewAI agent from a LangGraph workflow. Ollama runs local LLM inference, eliminating API keys and reducing latency. The guide walks through setting up the environment, installing dependencies, configuring LangGraph, MCP, A2A, and Ollama, and building the four agents step by step. It also covers state persistence with SQLite checkpoints, human oversight hooks, and observability with Langfuse and DeepEval for automated quality checks. The final system is a learning accelerator that plans study roadmaps, explains topics, runs quizzes, and adapts based on results. The handbook provides code, architecture diagrams, and best‑practice checklists for deploying similar multi‑agent systems in production.
Key changes
- Use LangGraph 1.1.0 for stateful orchestration.
- MCP 1.26.0 for standardized tool access.
- A2A SDK 0.3.25 for cross‑framework agent coordination.
- Ollama for local LLM inference (qwen2.5 models).
- Langfuse 4.0.1 for distributed tracing.
- DeepEval 3.9.1 for automated quality checks.
- SQLite checkpoint store for state persistence.
- Human approval node for oversight.