langgraph

Define a TypedDict state with an add_messages reducer, bind tools to an LLM, create an agent node and a ToolNode, then add a conditional edge that routes to tools when tool calls exist and loops back to the agent. Compile the StateGraph and invoke it with user messages.

Reducers control how node outputs merge into state. Annotated fields like Annotated[list, add_messages] append instead of overwrite, while custom functions can merge dicts or accumulate counters. Nodes return partial updates and LangGraph applies the reducers.

Yes, LangGraph supports checkpointers that persist graph state between runs. This enables resuming conversations, human-in-the-loop interruptions, and durable execution for production agents.

Add explicit exit conditions in your routing function, such as a maximum iteration counter stored in state or a task-complete flag that returns END. Also enforce timeouts at the application level to stop runaway cycles.

LangGraph is primarily Python-only, with TypeScript support in early stages. It has a learning curve around graph concepts, state management adds complexity, and debugging cyclic graphs can be challenging.