LangGraph’s opinionated older sibling for long-horizon agents

What it does

Deep Agents is an opinionated agent harness sitting atop LangGraph and LangChain. It gives you an agent that can plan across long horizons, delegate to sub-agents with isolated context, read and write files, manage its own memory, and call tools—including MCP servers—out of the box. You can swap models, prompts, or entire sub-components without forking the codebase.

The interesting bit

The project is explicitly inspired by Claude Code but generalizes the pattern: instead of baking in one model or sandbox, it treats planning, context summarization, and human-in-the-loop approval as pluggable defaults. The security stance is unusually candid—it assumes you must sandbox the tools, because the LLM itself is trusted to use them.

Key highlights

• Ships with sub-agents, filesystem access, shell execution, and persistent memory as first-class features
• Model-agnostic by design: works with any LLM that supports tool calling, from frontier APIs to local Ollama instances
• Built on LangGraph, so it inherits streaming, checkpointing, and persistence; pairs with LangSmith for tracing
• Lets you override or replace any component without forking, and can drop down to raw LangGraph graphs when the default loop is too rigid
• Also available as a JavaScript/TypeScript library

Caveats

• The “trust the LLM” security model means you are responsible for sandboxing tools; the agent will not self-police
• It is a layered abstraction: if you need a custom graph shape, you may still need to drop down to LangGraph directly

Verdict

Use this if you want a ready-to-run agent with planning, memory, and delegation already integrated. Skip it if you prefer to hand-roll your own agent loop or need a graph architecture that diverges sharply from the default harness.