How to Start Using LangGraph

LangGraph is a Python framework for building stateful agents as graphs (nodes + edges) or with a Functional API—part of the LangChain ecosystem. Start from the docs: docs.langchain.com (LangGraph quickstarts live under the LangChain docs). Product overview: langchain.com/langgraph.

For after-work Python, LangGraph teaches the important idea: an agent is a loop with state, not a single chat completion. Begin with a calculator-style tool agent—not a multi-service RAG empire.

Who this is for

Learners who already know basic Python functions/classes and want their first programmable agent loop. If you have never written a function with type hints, spend a few Intro lessons first.

Overview: Top 10 AI agent tools for after-work Python.

Steps to start tonight

1. Create a tiny venv and install

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -U langgraph langchain langchain-anthropic
# or your preferred chat-model integration from the docs

Follow the current LangGraph quickstart for the exact package names your doc version recommends.

2. Set one API key in the environment

export ANTHROPIC_API_KEY=...   # or OPENAI_API_KEY / provider of choice

Never paste keys into notebooks you will commit. Prefer a local .env that stays gitignored.

3. Copy the smallest official graph example

Use the docs’ Graph API or Functional API calculator agent (add/multiply/divide tools). Prefer the shortest sample that compiles and invokes—delete visualization helpers if they distract.

4. Run one HumanMessage and print the trail

Invoke with something like “Add 3 and 4.” Print each message. Confirm a tool call happened, then a final answer. That trail is the lesson.

5. Save NOTES.md: state vs tools vs edges

Write in your own words: what lived in state, which node called tools, when the graph ended. This note beats bookmarking ten blog posts.

Tiny exercise (20–30 minutes)

Add a subtract(a, b) tool, rebind tools, re-run “Subtract 10 and 3.” If the model ignores the new tool, improve the tool docstring—do not rebuild the whole graph.

Pitfalls

• Installing half the LangChain monorepo. Start with what the quickstart lists; expand later.
• Skipping state. If you only call model.invoke, you are not learning LangGraph yet.
• Mega RAG on night one. Documents + vector DBs can wait until the loop is clear.
• Hard-coding secrets. Keys in source will haunt you.
• Chasing every API rename. Pin docs version; APIs evolve—copy from current quickstart.

Why this fits after work

One venv, one graph, one arithmetic task fits a weekday evening. You leave with a mental model of agent loops you can reuse in other frameworks.

Next step

When tool-calling graphs feel obvious, strengthen Python foundations with Intro to Python Programming so framework samples stop feeling like magic.