framework-selection

Choose based on task complexity: LangChain for single-purpose agents and chains, LangGraph for custom control flow with loops and branching, and Deep Agents for open-ended tasks needing planning, file management, and subagent delegation.

Use LangGraph when the agent needs branching logic, loops, parallel fan-out, human-in-the-loop approval, or persistent state across invocations. LangChain suits fixed tool loops, RAG pipelines, and simple model calls.

Deep Agents handles long-running, multi-dimensional tasks with built-in middleware for todo lists, filesystem tools, subagent delegation, on-demand skills, and persistent memory across sessions.

Yes, a LangGraph compiled graph can be registered as a subagent inside Deep Agents. The Deep Agents orchestrator delegates to it via the task tool without needing to know its internal graph structure.

No, LangGraph has no middleware layer; behavior is wired directly into nodes and edges. Explicit configurable middleware such as TodoListMiddleware and FilesystemMiddleware is specific to Deep Agents.

Avoid Deep Agents when the task is simple enough for a single-purpose LangChain agent, or when you need precise hand-crafted control over every graph edge, in which case LangGraph is the better fit.