Build an AI Agent from Scratch with LangGraph

Most AI agent tutorials start with a framework and a magic create_agent() call. When the agent loops forever, leaks data or answers with confidence from the wrong document, you are left guessing what happened inside.

In this course you build the agent yourself first. You write the agent loop in plain Python: a short loop that calls a model, runs tools and feeds the results back. Then you rebuild the same agent with LangGraph, so every node, edge and checkpoint has something you already understand behind it.

Every chapter has two parts:

- From scratch: a small Python package, tiny_agents, that grows with the course: tools and function calling, the ReAct loop, structured answers, RAG, memory, planning, a code sandbox, multi-agent teams, evaluation, guardrails and tracing. - The LangGraph version: the same idea as a graph: state and reducers, ToolNode, conditional edges and Send, agentic RAG, checkpointers and long-term memory, plan-and-execute, human-in-the-loop with interrupt(), supervisors, handoffs and subgraphs, testing, retries, caching and durable execution.

What makes it different

- Real runs, honest numbers. Every result in the videos comes from a real model call, recorded and replayable. When the model skips an instruction, falls for a prompt injection or loops for eight calls, the course shows it, explains why from the trace, and fixes it. - Free to follow. Examples run on open-weight gpt-oss models through Groq's free tier, or on OpenAI. The last chapter runs the agent on small open models on a laptop with Ollama. - The 2026 stack. MCP tools, the A2A protocol, the OpenAI Agents SDK and PydanticAI compared with LangGraph, prompt-injection defences, cost per answer, and a FastAPI service in a Docker container.

What is inside

- 15 sections, 64 short lectures, about 1 hour 50 minutes of focused video. Every number on screen comes from the notebooks. - 15 quizzes with 194 questions. - 27 hands-on Udemy Labs: exercises with checks that print ✓ when your code is right. - Practice and solution notebooks for every lesson, also as self-contained Google Colab notebooks. - English narration with English and Vietnamese captions.

Course outline

1. What is an AI agent? · LangGraph 101: state, nodes and edges
2. LLM APIs for agents · reducers and streaming
3. Tools, function calling and MCP · ToolNode and MCP tools
4. A ReAct agent from scratch · the ReAct graph
5. Workflows vs agents: six patterns · conditional edges and Send
6. Knowledge with RAG and file tools · agentic RAG
7. Memory that lasts · checkpointers, threads and long-term memory
8. Planning and reflection · plan-and-execute with a critic
9. Running code safely · human-in-the-loop with interrupt()
10. Multi-agent systems and A2A · supervisors, handoffs and subgraphs
11. Evaluating agents · testing and evaluating graphs
12. Agent frameworks in 2026 · LangGraph in production
13. Security and guardrails
14. Observability and cost
15. Open models and deployment

By the end you will not just know how to call an agent framework. You will know what it does for you, and you will be able to debug it when it does not.

Narration uses an AI-synthesized clone of the instructor's own voice. Code builds on the MIT-licensed shangrilar/ai-agent-from-scratch repository. LangGraph, the OpenAI Agents SDK and PydanticAI are open-source projects by their authors. Examples, animations, exercises and quizzes are original to this course.