Deep Agents is a modular agent architecture in the LangChain ecosystem that combines planning, long-term filesystem-backed memory, and the ability to spawn subagents to solve complex, multi-step problems. Implemented as middleware, Deep Agents automatically attach PlanningMiddleware, FilesystemMiddleware, and SubAgentMiddleware when created via create_deep_agent, enabling structured task decomposition, persistent context storage, and delegation. The system supports human-in-the-loop workflows through LangGraph interrupts and a checkpointer/HITL middleware for tool approval, and the resulting agent is a LangGraph graph that can be interacted with via streaming, memory APIs, and studio tooling. Unique value comes from its composable middleware design which lets developers build research-grade, long-horizon agent workflows that balance automation, memory persistence, and human oversight.