What's the difference between the Claude Agent SDK and Deep Agents?

The Claude Agent SDK is a single-model library, embedding Claude Code’s own agent loop, tools, and permissions into a process you run yourself; Deep Agents is a standalone library built on LangChain’s own agent-building blocks, using the LangGraph runtime underneath for durable execution, that reimplements a similar shape, a todo-list planner, a virtual filesystem, and subagents, for whatever model you wire in through LangChain’s connectors. The two converge on similar structure for different reasons: Claude Code’s subagents and file tools are native to how Claude itself works, while Deep Agents builds an equivalent planning tool and virtual filesystem in code to give that same structure to a model-agnostic loop.

That difference shows up in more than branding: Deep Agents can run on any model LangChain’s connectors reach, while the Claude Agent SDK only runs Claude. Deep Agents also inherits durable execution from the LangGraph runtime underneath: LangGraph is what actually checkpoints and resumes a run, not something Deep Agents adds on top. The Claude Agent SDK checkpoints nothing inside a single session, only sessions that resume or fork as a whole.