LangGraph
A stateful agent runtime published by LangChain Inc, usable without the LangChain components. It decides how steps run: in a line, in a loop, with state that survives from one step to the next, conditional routing and cross-turn persistence. The example that keeps coming up in interviews (“a research agent is asked: give me information about Apple stock”): with a chain, the Google search, then the database, then the LLM run one after another, and if the search fails the whole path can fail. With LangGraph the agent picks the next step at runtime from the result it just received: call an API, query the database or retry the search. The steps are called nodes, the links between them edges. Source: [Instagram reel DeOtQ_MMdS-](https://www.instagram.com/reel/DeOtQ_MMdS-/).
Strengths
- Cyclic graphs: agents can loop, re-evaluate, and change strategy mid-run
- Built-in persistence: cross-turn memory, checkpoints, state snapshots
- Native human-in-the-loop: pause agent for human validation, then resume
- Native streaming: each graph node streams its output in real-time
- LangGraph Platform: persistent agent deployment in production (hosted)
- Usable without LangChain: the official documentation says so, you plug your own model and tool calls into the runtime
Limitations
- Increased complexity vs simple chains, debugging cyclic graphs is non-trivial
- LangGraph Platform (hosted) is paid, self-hosting requires infra work
- Overkill for simple agents (question → tool call → response)
Best for
- Multi-step agents with memory, loops, and human validation
- Complex LLM workflows: iterative search, planning, self-correction
- Enterprise chatbots requiring persistent state across sessions