What's the difference between LangGraph and Deep Agents?

LangGraph is the low-level runtime, nodes as functions, edges deciding what runs next, one state object threaded through the whole run; Deep Agents is a batteries-included agent built on top of that runtime, shipping a planning tool, a virtual filesystem, and subagent spawning already wired in rather than left for you to write. LangChain’s own comparison calls Deep Agents the architecture behind tools like Claude Code and Manus: an agent that plans a todo list, works through a virtual filesystem instead of passing every file through its context, and hands pieces of the task to subagents that run and report back.

Because Deep Agents runs on the identical LangGraph engine underneath, checkpointing and replay work the same way: a thread id still saves state after every step, and you can rewind to one and resume it. What changes is what there is to grade. A plain LangGraph agent gives you the nodes you wrote; a Deep Agents run adds nodes you didn’t, the planning step and each subagent call, so judging one subagent apart from the run that launched it becomes a real question the moment you adopt it instead of writing your own graph by hand.