# Deep Agents

An open-source, opinionated agent harness built on LangGraph, with sub-agents, a pluggable filesystem, persistent memory, human-in-the-loop approval and loadable skills bundled in by default.

Deep Agents runs out of the box with defaults tuned for long-horizon, multi-step work, while letting any piece be overridden, extended or replaced; it works with any LLM that supports tool calling, whether a frontier API, an open-weight model, or a local one. Bundled capabilities include delegating tasks to sub-agents with isolated context windows, a filesystem abstraction for reading, writing, editing and searching content across local, sandboxed or remote backends, summarizing long conversation threads and offloading tool output to disk, shell access inside a chosen sandbox, pluggable persistent memory for cross-session recall, human-in-the-loop approval of tool calls before they run, and loadable skills and custom tools including any MCP server.

It is built on LangGraph for streaming, persistence and checkpointing, with tracing, evaluation and deployment through LangSmith, and is also available as a JavaScript/TypeScript library. A separate pre-built terminal coding agent, Deep Agents Code, is built on the same harness.

It is written in Python and licensed under the MIT license.

Website: https://docs.langchain.com/deepagents