langgraph-fundamentals

Define a state schema as a TypedDict, add node functions that return partial state updates, connect them with add_edge or add_conditional_edges, then call compile() before invoking. The graph must be compiled to produce an executable object.

List fields without a reducer are overwritten by each node update, so earlier values are lost. Fix it by annotating the field with a reducer, such as Annotated[list, operator.add] in Python or ReducedValue in TypeScript, so updates accumulate.

Command combines a state update and a single routing decision in one node return value. Send fans out to multiple parallel workers from a conditional edge, and requires a reducer on the results field to aggregate worker outputs.

Use LangGraph when you need fine-grained control over orchestration, complex branching or loops, human-in-the-loop, or persistence. For quick prototyping or simple stateless workflows, LangChain agents or direct calls are a better fit.

Attach a RetryPolicy with max_attempts to nodes for transient failures like network errors. For tool failures, use ToolNode with handle_tool_errors enabled so errors return as ToolMessages the LLM can recover from.

Static edges that point back between two nodes create an infinite cycle. Add a conditional edge that returns END once a termination condition is met, such as a counter exceeding a threshold.