Deep Agents
An open-source AI agent framework handles multi-step work with task delegation, file access, and support for local models through Ollama, vLLM, or llama.cpp.
Deep Agents is an AI agent framework for developers building assistants that carry out multi-step work. It provides planning and task execution defaults that you can customize without forking the project.
- Planning and delegation Agents can plan work and assign tasks to sub-agents with separate contexts.
- Files and commands Agents can read, edit, and search files across local, sandboxed, or remote storage, and run shell commands in a chosen sandbox.
- Context and memory The framework summarizes long conversations, moves large tool outputs to disk, and supports memory across sessions.
- Tools and skills You can connect your own functions or MCP servers and give agents reusable behaviors they load when needed.
- Approval controls People can approve, edit, or reject tool calls before execution.
The MIT-licensed framework is open source, with Python and JavaScript/TypeScript libraries. It works with tool-calling models through LangChain, including self-hosted models served by Ollama, vLLM, or llama.cpp and cloud APIs from OpenAI, Anthropic, or Google. Choosing a local model doesn't make every connected tool local; file storage and external services depend on the backends you use.
It builds on LangGraph for streaming, saved state, and checkpoints. LangSmith provides optional tracing, evaluation, monitoring, and deployment. Deep Agents Code is a ready-made terminal coding agent built on the framework. Agents can perform any action their tools permit, so access restrictions must live in the tools or sandbox.
Logotyper
