Deep Agents

The page presents deepagents as a Python package for building agents that can handle longer, more complex tasks. It says the package combines a planning tool, subagents, access to a file system, and a detailed prompt to move beyond shallow tool-calling loops.

Reading notes

#

• deepagents is described as a general-purpose way to create a Deep Agent for an application.
• The project says it was primarily inspired by Claude Code.
• The example shows create_deep_agent being used with a search tool and research instructions.
• A deep agent created with create_deep_agent is said to be a LangGraph graph, so it can use streaming, human-in-the-loop, memory, and studio.
• create_deep_agent takes tools and instructions as required parameters.
• It also accepts optional subagents, each defined with a name, description, prompt, and optional tools.
• By default, deepagents uses the model claude-sonnet-4-20250514.
• The built-in system prompt is described as detailed and as necessary for the agent to work well on deep tasks.
• The built-in planning tool is compared to Claude Code’s TodoWrite tool and is used to keep a plan in context.
• The file system tools are ls, edit_file, read_file, and write_file, and they are mocked through LangGraph state rather than a real file system.
• The file system is currently one level deep and uses the files key in the LangGraph State object.
• The package includes a built-in general-purpose subagent and allows custom subagents for context quarantine and custom instructions.
• The roadmap includes custom system prompts, cleaner code, a more robust virtual filesystem, a deep coding agent example, benchmarking the deep research agent, and human-in-the-loop tool support.