This Python package, deepagents, enables the creation of advanced AI agents capable of complex, multi-step tasks by integrating planning, sub-agents, and file system interaction. It targets developers and researchers seeking to build agents that go beyond simple tool-calling loops, offering enhanced capabilities for applications like in-depth research and code generation.

How It Works

The core of deepagents is built upon LangGraph, allowing for the creation of stateful, multi-agent workflows. It leverages a detailed, pre-defined system prompt inspired by Claude Code, which guides the agent's behavior. Key features include a built-in planning tool for task decomposition, mocked file system tools (read, write, edit) for managing state without actual disk I/O, and a flexible sub-agent architecture for specialized tasks or context management.

Quick Start & Requirements

• Install: pip install deepagents
• Prerequisites: TAVILY_API_KEY environment variable for the internet_search tool. Optional: langchain, langchain-ollama for custom models, langchain-mcp-adapters for MCP tools.
• Example usage and more complex research agent examples are available in the repository.

Highlighted Details

• Implements a "deep" agent architecture inspired by Claude Code, focusing on planning and multi-step execution.
• Provides built-in, mocked file system tools for state management within LangGraph.
• Supports custom sub-agents with distinct instructions and tool access for modularity.
• Allows customization of the underlying LLM, including local models via Ollama.
• Can integrate with Multi-Server MCP tools for extended capabilities.

Maintenance & Community

The project is actively maintained by hwchase17. A roadmap is provided, outlining planned improvements such as full system prompt customization, code refinement, enhanced virtual filesystem, and benchmarking.

Licensing & Compatibility

The project's license is not explicitly stated in the README. Compatibility for commercial use or closed-source linking would require clarification on the licensing terms.

Limitations & Caveats

The current virtual file system is one level deep and does not support subdirectories. The roadmap indicates that benchmarking and human-in-the-loop support are planned features, suggesting they are not yet implemented.