LangGraph¶
5 weeks on LLM work as a state machine: nodes are functions, edges are control flow, state is the request-scoped object you already thread through a saga.
Start this track only if state machines, retries, persistence, idempotency, and the course prerequisites are familiar.
Scope: get the state-machine idea, build a resumable graph, then use the official docs to go further. Read the framework track guide before starting.
Use a graph when you need branches, retries, pause-for-a-human, or replay. Three sequential calls do not need a graph.
| Week | Idea | You are done when… |
|---|---|---|
| 1 — Graphs | StateGraph, reducers, a graph that branches. |
Two inputs take different visible paths through the graph. |
| 2 — Workflows | Conditional edges, fan-out, subgraphs. | Parallel branches join into one validated state. |
| 3 — Persistence | Checkpoints. Replay after a crash. | A failed run resumes without repeating completed pure work. |
| 4 — Human in the loop | Interrupt, resume, the approval node. | Approve, reject, and needs-info follow distinct tested paths. |
| 5 — Don’t charge twice | Resume is at-least-once. Keys make effects once. | Replaying the side-effect node produces one charge. |
| Capstone — Support agent | LangChain 4/5/7 + LangGraph 3/4/5 in one graph. | Six golden tickets pass, and the gate fails an agent without the approval interrupt. |
1 Graphs ──▶ 2 Workflows ──▶ 3 Checkpoints ──▶ 4 Approval ──▶ 5 One charge
│
▼
Support agent (also LangChain 4, 5, and 7)
The support agent is the ticket bot that can move money. Week 17 is the night the churn list breaks. The ML on-call specialist is a read-only assistant for that runbook, and it comes after the job-path capstone.
Think of it like… CI, Redux, or a workflow engine
The runtime owns the state so you can stop in the middle and start again. That is the product feature. Pretty diagrams are a side effect.