langgraph-fundamentals

Define a state schema, add node functions that return partial state updates, connect them with add_edge or add_conditional_edges, then call compile() before invoking. Follow the five-step design process: map steps, identify node types, design state, build nodes, and wire edges.

Use LangGraph when you need fine-grained control over orchestration, complex workflows with branching or loops, or human-in-the-loop and persistence features. Use LangChain agents for quick prototyping and simple stateless workflows, or Deep Agents for batteries-included features.

List fields without a reducer are overwritten by each node update, so only the last write survives. In Python use Annotated[list, operator.add]; in TypeScript use ReducedValue with a concat reducer to accumulate values across nodes.

Call graph.stream with stream_mode set to "messages" to receive LLM tokens plus metadata as they are generated. Other modes include values for full state, updates for state deltas, and custom for user-defined progress data emitted via a stream writer.

Command combines a state update and a goto routing decision in a single node return value. Send is returned from a conditional edge to fan out to multiple parallel workers with dynamic inputs, and requires a reducer on the results field to aggregate outputs.

The graph builder is not executable until you call compile(). Invoke the compiled graph returned by builder.compile(), not the builder itself, and in TypeScript remember to await the invoke call since it returns a Promise.