LangGraph

LangGraph is LangChain's low-level runtime for building AI agents as state graphs. Developers write nodes and edges in Python or JavaScript, and the runtime runs them step by step. A checkpointer such as Postgres or SQLite saves state for each thread, so long-running runs can resume after a failure or a pause. Built-in interrupts let a person approve or edit an agent's action before it continues. The framework also supports token streaming, persistent memory, and single-agent, multi-agent, and hierarchical control flows. It works with any model vendor. It suits engineering teams building production agents that need explicit control flow, durability, and auditability. The vendor's site lists users such as Elastic, Cisco, Uber, Lyft, and Workday. The library itself is free. Costs come from model calls and your own database, or from LangSmith if you want hosting. The LangSmith Developer plan is free for one seat with 5,000 traces a month. Plus costs $39 per seat per month and adds the managed Agent Server, which bills compute and memory usage in LangSmith units. Enterprise has custom pricing and adds hybrid and self-hosted deployment, SSO, and RBAC. Tracing and observability come through LangSmith, which is a separate product. The main limitation is the learning curve. Teams design the agent loop themselves, so there is more upfront code than with ready-made agent builders. Thoughtworks moved it from Adopt back to Trial in April 2026 and no longer treats it as the default choice for every agentic system.