Deep Agents

LangChain distilled the architecture behind coding and deep-research agents into deepagents, a harness that upgrades any tool-calling LLM into a long-horizon worker. On top of the LangGraph runtime it layers the ingredients such agents need: a planning tool that maintains a running todo list, a virtual filesystem plus sandboxed shell for reading, writing, and executing work products, sub-agent delegation so focused workers burn context in isolation, and automatic summarization that keeps long sessions inside the context window. Human-in-the-loop approval gates, persistent cross-session memory, MCP server support, and pluggable skills round out the package. Installation is pip or uv add deepagents for Python, with a parallel deepagents.js package for TypeScript, and the model layer swaps freely across frontier APIs, open-weight models, and self-hosted endpoints. MIT licensed with about 28k GitHub stars and documentation on LangChain's docs site, it has become one of the most widely adopted agent scaffolds of 2026.