LangGraph

Agent loops that must survive crashes, pause for a person, or branch on intermediate results are what LangGraph is built for. You describe the application as a StateGraph: nodes are Python functions or model calls, edges, including conditional ones, decide what runs next, and a shared state object flows between them, a model the authors credit to Google's Pregel and Apache Beam. A checkpointer saves state after every step, so a run can resume where it stopped, be paused and edited mid-flight through interrupts, and keep per-thread working memory plus a long-term store across sessions. The repository contains the core library, prebuilt agent helpers, SQLite and Postgres checkpoint backends, Python and JavaScript client SDKs, and a CLI, while LangGraph.js covers TypeScript. It needs Python 3.10 or later and depends on langchain-core but not the full LangChain package. The library is MIT licensed and free for commercial use; hosted deployment through LangSmith is an optional paid product. Choose it when you need explicit control over agent flow, durability and approval steps rather than a high-level agent wrapper.