LangGraph
State-machine framework for LLM agents — define nodes (steps) and edges (transitions) and run them as a graph with explicit, checkpointable state.
From the LangChain team in 2024, a response to the fact that regular "chain" abstractions don't suffice when agents need cycles, branching and human-in-the-loop. State is a plain TypedDict; nodes are functions that mutate/return a state delta; edges can be conditional.
A checkpointer (Postgres, Redis) saves state after every node — you can pause, resume, time-travel and plug in human approval. The modern choice for production agents in 2025.