Deep Agents — готовая обвязка для AI-агентов

★ 29.8K

Deep Agents is an open-source Python library — a batteries-included agent harness from the LangChain team — for running long-horizon, multi-step AI agents with built-in filesystem access, context management, and sub-agent delegation, without assembling the plumbing from scratch. Reach for it when you need an agent that can plan across many steps, read and write files, run shell commands, and remember state across sessions — or when you want human-in-the-loop approval before any tool fires, or need to plug in an MCP server as a tool source. Under the hood: a task planner, sub-agents with isolated context windows, long-thread summarization, pluggable filesystem and persistent-memory backends, and human-in-the-loop tool-call approval. Built on LangGraph (streaming, checkpointing, persistence) and model-agnostic — works with any tool-calling LLM: OpenAI, Anthropic, Google frontier APIs, or self-hosted models via Ollama, vLLM, or llama.cpp. Install with `uv add deepagents`; a JavaScript/TypeScript version lives in the separate deepagentsjs repo. The library sits above LangGraph (the graph runtime) and above LangChain's lighter `create_agent` wrapper — use Deep Agents when you want the full middleware bundle out of the box, and drop to LangGraph directly when the agent loop itself needs a custom shape.