Deep Agents

Deep Agents is an agent harness built on LangGraph: an opinionated agent that runs without configuration for long-horizon, multi-step work, where any piece can be extended, overridden, or replaced without forking. It sits above LangChain's create_agent and bundles a filesystem, sub-agents, context management, and skills that the lighter harness leaves out.

You create an agent with createdeepagent, passing a model, tools, and a system prompt; the agent can plan, read and write files, and manage its own context. It works with any LLM that supports tool calling, whether a frontier API (OpenAI, Anthropic, Google), an open-weight model on a hosting provider, or a self-hosted model via Ollama, vLLM, or llama.cpp, through any LangChain chat model. Because it runs on LangGraph, streaming, persistence, and checkpointing come from the runtime, and LangSmith provides tracing, evaluation, and deployment.

The layers compose: any compiled LangGraph state graph can be passed in as a sub-agent, so custom orchestration plugs in alongside the harness defaults. A JavaScript/TypeScript port, deepagents.js, is maintained separately, and Deep Agents Code is a pre-built terminal coding agent on the same harness. The project follows a trust-the-LLM security model: the agent can do anything its tools allow, so boundaries belong at the tool and sandbox level.

Features

• Sub-agents: delegate tasks to agents with isolated context windows
• Filesystem: read, write, edit, and search over pluggable local, sandboxed, or remote backends
• Context management: summarise long threads and offload tool outputs to disk
• Shell access: run commands in a sandbox of your choice
• Persistent memory: pluggable state and store backends for cross-session recall
• Human-in-the-loop: approve, edit, or reject tool calls before they run
• Skills: reusable behaviours the agent loads on demand
• Tools: your own functions or any MCP server
• Model-agnostic: any tool-calling model through LangChain chat model integrations
• LangGraph runtime: streaming, persistence, and checkpointing, with LangSmith for tracing and evals