LearnLangGraph
A hands-on LangGraph course in three waves: graphs without an LLM, then agents, then production patterns.
Active · join meAI Engineering
1 · Purpose
LearnLangGraph teaches the LangGraph library by doing, one small runnable lesson at a time. Wave 1 uses no language model at all, so state, nodes, routing and loops are learned on plain code. Wave 2 adds a free open-weight model, and the agent loop becomes just one more node. It is for engineers who want to build agents on a clear mental model rather than copy examples.
2 · Business benefits
Faster onboardingA new engineer can go from zero to a tool-using agent with memory in a few evenings.
Fewer fragile agentsLearning routing and loops on deterministic code first makes later agent bugs easier to find.
Zero model spend while learningLessons run on a free model tier, and tests use a scripted fake model with no network.
A reusable teaching assetEvery lesson is a runnable file plus a test, so a team can use it as its own internal course.
3 · Status and how to join
Active · join me. Under way: pick a task below, or tell me what you'd like to learn and I'll suggest one.
Good first tasks this week:
- Write lesson 15: an agent proposes a fix and pauses for approve or reject.
- Add 'try this' answers to a Wave 2 lesson.
- Run the course on a fresh machine and fix any setup step that breaks.
4 · Requirements
What it must do
- Teach state, nodes, edges, conditional routing, reducers, parallel branches, loops and streaming with no LLM.
- Finish Wave 1 with a capstone graph that triages Spark job anti-patterns.
- Teach an LLM as a node, raw tool calling, a hand-built ReAct loop, the prebuilt agent, memory and structured-output routing.
- Finish Wave 2 with a capstone cost assistant that combines a guard, an agent, tools and memory.
- Plan Wave 3: human approval, time travel, subgraphs, map-reduce, supervisor multi-agent, durable storage and evals.
- Publish the whole course as one generated HTML page with diagrams, code, real output and exercises.
How well it must do it
- Every lesson has a table-driven test.
- Tests need no API key; a scripted fake model replaces the real one.
- Switching model provider is two settings, not a code change.
- Real model output is cached so the page can be rebuilt without calling the model.
5 · Architecture, very high level
6 · Tech stack, per platform
| Component | Open stack | Azure | AWS |
|---|---|---|---|
| Lesson files | Python 3.12 + LangGraph 1.x ✓ | Azure Container Apps for a hosted agent | AWS Lambda or ECS for a hosted agent |
| Shared tools | Plain Python ✓ | — | — |
| Model adapter | OpenAI-compatible client on Groq free tier ✓ | Azure OpenAI / Azure AI Foundry | Amazon Bedrock |
| Scripted fake model | Custom LangChain chat model ✓ | — | — |
| Tests | pytest, table-driven ✓ | — | — |
| Page builder | Python script + HTML template ✓ | — | — |
7 · What needs to be done
Wave 1: basics without an LLM 2 of 2 done
| Task | Status | Size | Skill |
|---|---|---|---|
| Lessons 01 to 07 with tests | done | M | Python |
| Spark triage capstone | done | S | Python |
Wave 2: agents 3 of 3 done
| Task | Status | Size | Skill |
|---|---|---|---|
| Lessons 08 to 14 on a free model | done | M | LLM agents |
| Offline fake model and cached real output | done | S | testing |
| Single-page HTML course | done | S | technical writing |
Wave 3: production patterns 0 of 5 done
| Task | Status | Size | Skill |
|---|---|---|---|
| Lesson 15: pause for human approval and resume | todo | S | LangGraph |
| Lesson 16: time travel and forking a run | todo | S | LangGraph |
| Lessons 17 to 18: subgraphs and map-reduce | todo | M | LangGraph |
| Lesson 19: supervisor with worker agents | todo | M | multi-agent |
| Lessons 20 to 22: durable storage, evals and capstone | todo | L | LangGraph |
8 · Roles
| Role | How many | Hours / week | What they do |
|---|---|---|---|
| Lesson author | 1 | 3-4 | Writes one Wave 3 lesson at a time with its test. |
| Tester / reviewer | 1 | 1-2 | Runs every lesson fresh and reports where it confuses. |
9 · What you'll learn
- A graph is a function: state goes in, state comes out.
- Reducers decide how two updates to the same key are merged.
- An agent is a graph with a loop where one node calls a model.
- Checkpointers give memory, pause and resume.
- How to test agent code without calling a model.
Level-up questions
- When does a graph framework help, and when is a plain loop clearer?
- How do you stop an agent loop safely?
- Where should a human approval step go in a real workflow?
10 · Links
- Links come once the work is public.