langgraph-fundamentals

You orchestrate multi-node agent workflows in LangGraph by constructing a StateGraph, defining nodes that mutate state, and connecting them with linear or conditional edges. This models execution order as a directed graph to reduce coordination errors.

State updates in a graph workflow are managed by defining typed State schemas and applying reducers. Nodes mutate state in a controlled, partial-update manner, ensuring data flows predictably through the directed graph.

LangGraph state machines support conditional routing and fan-out via Send. You can define edge types including linear, command-based routing, and fan-out to direct execution flow dynamically based on state.

LangGraph supports streaming and debugging for LLM-integration pipelines through built-in commands and example-driven guidance. It handles streaming outputs and provides strategies for error handling during execution.

To compile and invoke a StateGraph, define a State schema, create nodes, connect START to the first node and end to END, then compile the graph and invoke it with an input. Examples are provided in Python and TypeScript.

LangGraph should be used for workflow orchestration instead of basic LLM chains when you need complex, multi-node agent pipelines. It provides clear execution order, state management, and advanced patterns like reducers for compliance checks and decision logging.