DeepAgents

A powerful agent harness for building autonomous AI systems with recursive planning, sub-agent delegation, and persistent file-system memory.

DeepAgents is the specialized framework for building autonomous AI systems that handle complex, multi-step workflows. Built on LangChain and LangGraph, the MIT-licensed Python library moves beyond simple tool-calling loops by providing a robust agent harness with recursive planning and sub-agent delegation. It uses a virtual file system to maintain persistent state across tasks spanning hundreds of steps (effectively preventing the context window overflow that plagues standard models). You can use built-in tools like write_todos to decompose high-level goals into discrete actions, while specialized sub-agents handle focused research or coding tasks. This architecture mirrors the design patterns found in production-grade systems like Claude Code and OpenAI's Deep Research, making it the standard for enterprise-level agentic automation.

https://github.com/langchain-ai/deepagents