LangGraph — State Machines for Agents

#The Problem

You ship a function-calling agent. It works for three turns, then something goes wrong: the model tries a tool that returns 500, the user changes their mind mid-task, or the agent decides to refund an order without a human signing off. The while True: loop has no hooks. You can't pause it, you can't rewind it, and you can't branch off into "what if the model had picked the other tool." The moment you ship this past a demo, the agent becomes a black box that either worked or didn't.

The next step is obvious once you see it. The agent is already a state machine — system prompt plus message history plus pending tool calls plus the next action. Make the state machine explicit: nodes for "the model thinks," "a tool runs," "a human approves," and edges for the conditional transitions between them. Once the graph is explicit, the harness gets four things for free: checkpointing (save state between steps), interrupts (pause for a human), streaming (stream tokens and intermediate events), and time-travel (rewind to a prior state and try a different branch).

LangGraph is the library that ships this abstraction. It is not an agent framework in the LangChain sense ("here is an AgentExecutor, good luck"). It is a graph runtime with first-class state, first-class persistence, and first-class interrupts. The agent loop is something you draw, not something you hand-write.