langgraph-fundamentals

To build a stateful LangGraph workflow, define a typed state schema, construct a StateGraph with nodes and edges, and compile the graph before invoking it to ensure your orchestrated LLM workflow executes reliably.

LangGraph state overwrites during parallel execution when you lack proper reducers. Defining reducer-backed state schemas for lists and accumulators ensures parallel Send operations aggregate results correctly instead of overwriting previous node outputs.

To update state in a LangGraph node, return partial update dictionaries from your node functions. Avoid mutating and returning full state objects directly, as this prevents compilation errors and maintains correct graph execution.

You need to use Command and Send semantics in LangGraph when implementing branching, fan-out aggregation, and dynamic routing. They allow you to direct graph execution without conflicting static edges, enabling complex multi-step workflows.

Yes, you can stream intermediate state and message tokens in a LangGraph workflow. The framework supports streaming over token outputs and progress updates while safely aggregating node outputs for later synthesis.

The correct way to structure a multi-step routing graph in LangGraph is to define a StateGraph with typed state, use reducers for list accumulation, and apply Command or goto semantics to route queries to the right branch without static edge conflicts.