langgraph-fundamentals

LangGraph enables building stateful, multi-actor applications using cyclic graphs. Define a StateGraph with TypedDict or Pydantic state, add nodes representing actors or logic steps, connect them with edges, select reducers to merge state updates, then compile and invoke the graph with initial state to orchestrate execution.

Override mode replaces the entire state field, while reducer mode merges updates into existing state. LangGraph provides built-in reducers like operator.add and add_messages to accumulate values; you define which mode applies per field in your state schema using TypedDict or Pydantic.

LangGraph supports conditional edges that route based on state values and normal edges for fixed paths. Fan-out/fan-in patterns run multiple nodes in parallel and merge results using configured reducers, enabling complex orchestration workflows within a single compiled graph.

Yes. LangGraph graphs support both synchronous and asynchronous invocation with optional streaming to observe intermediate state updates. Persistence is available through the runtime execution layer, allowing you to checkpoint and resume graph execution.

LangGraph accepts TypedDict or Pydantic for Python state schemas, and Zod for TypeScript. Both approaches let you define typed fields, assign reducers to handle updates, and use override semantics where needed, ensuring type safety and clear state evolution.

No. LangGraph abstracts graph concepts into nodes (computation steps) and edges (connections), compiled into an executable workflow. You define state, add logic nodes, connect them, and invoke—no formal graph theory required to build working multi-step applications.