LangGraph#

LangGraph models agent workflows as state machines (graphs).

Why Graphs?#

Standard LLM chains are linear. Agents require loops (e.g., trying a tool, failing, thinking, trying again). Graphs natively support cyclic execution.

Core Concepts#

• State: A typed object passed between nodes.
• Nodes: Python functions that read the state, do work, and return state updates.
• Edges: Conditional routing logic (e.g., “If tool failed, go to ErrorNode”).

Human-in-the-Loop & Checkpointing#

LangGraph can persist state to a database. You can pause execution, wait for a human to approve an action, and resume exactly where it left off.