langchain-ai/langgraph

LangGraph is a low-level framework for building stateful, long-running AI agents and workflows. The key problem it addresses is that AI agents built from simple LLM calls are stateless and fragile: if a multi-step task fails partway through, you lose the progress, and there is no easy way to pause for human input or inspect what happened. LangGraph provides the infrastructure to make agents durable and observable. The core concept is representing an agent or workflow as a graph where nodes are computational steps (calling an LLM, running a tool, executing code) and edges define how control flows between them. State is persisted at each step, so if an agent crashes at step seven, it can resume from step seven. This durable execution is built in, as is the ability to pause a workflow at any point for a human to review or modify the agent's state before continuing. LangGraph also provides memory management: short-term working memory within a session and long-term memory that persists across sessions. This allows agents to accumulate knowledge and context over time. Integration with LangSmith provides visibility into what each agent step did, making it easier to debug complex multi-step agent behavior. You would use LangGraph when building production AI agents that need to run for extended periods, coordinate multiple tools or sub-agents, require human oversight at checkpoints, or need to be reliably restartable after failures. It is used for customer service automation, code generation pipelines, document analysis workflows, and other complex agent systems. The primary language is Python, and a TypeScript version called LangGraph.js is available separately. LangGraph can be used independently or alongside the LangChain library and the LangSmith deployment and observability platform.