Deep Agents Code

About

deepagents exists because teams kept rebuilding the same scaffolding around LLM agents: a planning step, file access, context compaction, and task delegation. The library ships those as defaults inspired by Claude Code — a built-in planning tool, pluggable filesystem (local, sandboxed, or remote backends), sub-agents with isolated context windows, shell access, persistent memory, and human-in-the-loop approval gates — while every component remains replaceable for teams with different needs. It runs on any tool-calling LLM through LangGraph's production features (streaming, checkpointing, persistence) and accepts any MCP server as a tool source. Python and JavaScript teams use it as the foundation for custom agents rather than as an end-user product, with LangSmith available for tracing and evaluation.

Other entries by langchain-ai

• Open SWE — agent
• LangChain JS Tools — agent-sdk