LangGraph

An agent loop written as a while-statement is easy to start and hard to live with: you cannot pause it, you cannot see why it took the branch it took, and a crash loses the run. LangGraph takes the loop out of the code and makes it a graph the runtime can inspect, interrupt and resume.

The model

• Nodes are functions — a model call, a tool execution, a validation step — each receiving the current state and returning an update to it.
• Edges connect them, and conditional edges route on the state, so cycles, branching and retries are expressed rather than buried in a loop body.
• State is one typed object with explicit rules for how each update merges in — the agent's memory for the run, not a growing prompt string.
• Checkpointers persist that state after every step, so a run survives a restart and can be rewound to an earlier step and replayed.

What persistence unlocks
Durable state makes the hard features ordinary. An interrupt suspends the graph before a sensitive tool call and waits hours for a human verdict, because nothing is held in memory. A failed run resumes from the last good step instead of starting over. Multi-agent setups become subgraphs under a supervisor node. And because every step is a recorded transition, a trace in LangSmith shows precisely where a run went wrong. The cost is explicitness: you draw the control flow up front, which is more work than a while-loop and the reason the graph earns its keep past a prototype.