deep-agents-core

To build deep agents with LangGraph, use an opinionated harness that provides middleware for task breakdown, context management, and persistent storage. Configure a model, tools, and a thread_id to maintain conversation context across sessions.

Human-in-the-loop approval for sensitive actions requires configuring a checkpointer. This safety workflow intercepts operations, allowing human oversight before execution within the agent's multi-step workflow.

Middleware handles task planning and subagent delegation by using TodoListMiddleware to break down tasks and delegating execution to subagents. This enables structured multi-step workflows without manual wiring.

Yes, you can configure a filesystem backend for on-demand skill loading in LangChain. This allows the deep agent to dynamically load and manage file context during multi-step workflows.

Yes, a checkpointer is required to maintain persistent memory across agent sessions. Configurable backends and thread-based configuration ensure conversation context is preserved between interactions.

Without an opinionated harness, building multi-step deep agents requires manually wiring task planning, tool use, context, delegation, and persistent memory. This increases development complexity and potential for integration errors.