Why Graphs?
From loops to LangGraph.
Our loop works. But the boss asks for more. Pause for approval on big refunds. Resume after a server restart. Roll back to an earlier step. Send one part down a different path. Suddenly, our twenty lines look like two thousand lines of messy code. That's the problem that graphs solve.
- Pause for approval
- Resume after a restart
- Twenty lines become two thousand
An agent as a graph
Think of your agent as a flow chart. Each box is a step, called a node. Each arrow is a connection, called an edge. Some arrows depend on a decision. And a shared notebook, called the state, travels along with the work. That's LangGraph, and the loop is just an arrow that points back.
Loop versus graph
Why not stick with a loop? A plain loop is great for simple agents, and you should keep using them. A graph pays off when you need to pause, resume, branch, or combine several agents. You're not giving up control. In fact, you can see the entire flow, and each transition is explicit.
Plain loop
- Simple to write
- Hard to pause or resume
- Flow hidden in code
Graph
- More to set up
- Pause, resume, roll back
- Flow is explicit
The three concepts
There are just three concepts. State is a typed dictionary holding everything that the agent knows. A node is a Python function that reads the state, and returns an update. And an edge decides where to go next. You already know all three of them. Our messages list was state, and our if statement was an edge.
Summary
State, nodes, and edges. That's the whole framework. Next, you'll build your first graph.
Key takeaways- Graphs make flow explicit
- Use them for pause and resume
- State, nodes, edges