DeepAgents: How LangChain's New Framework Gets AI Actually Working

Forum topic · 小凯 · 2026-03-23

Summary

Most AI assistants today act like sophisticated search-plus-text-generation tools: they offer advice and outlines but do not execute tasks. DeepAgents, a new open-source framework from the LangChain team, aims to change this by turning AI from a talker into a genuine task-performing assistant. The framework is built on four pillars: (1) a planning tool that automatically generates to-do lists and tracks progress; (2) a full filesystem layer with tools like create_file, read_file, edit_file, and ls, enabling the agent to organize and modify files; (3) sub-agents spawned via a task tool, each with an independent context to handle complex subtasks without interference; and (4) context management that summarizes long conversations and offloads large outputs to files to combat context loss. Inspired by Anthropic's Claude Code, DeepAgents differs by supporting any LLM rather than Claude only, being MIT-licensed open source, and being LangGraph-native with LangSmith, persistence, and checkpointing support. A one-line constructor creates an agent with all features enabled out of the box.

Have you ever wondered this question:

Today's AI can chat, write articles, even help you code — but can it independently complete a task?

Say: "Research the EV market for me and write a report." The AI returns a long reply about what to research, where to get data, and how to structure the report.

Sounds professional, right? The problem is — you still have to do the actual work yourself.

No surprise there. Most AI assistants are essentially "a fancy search engine + text generator": they advise but don't execute; they inform but don't act. Like a knowledgeable consultant who points from the sidelines but never rolls up their sleeves.

Enter DeepAgents, a new framework from the LangChain team. Its ambition is simple: turn AI from a parrot that only talks into an assistant that actually works.

🛠️ Four Pillars: DeepAgents' Core Weapons

📝 Pillar 1: Planning — the art of fitting an elephant into a fridge

DeepAgents ships with a built-in planning tool. When you assign a task, it automatically generates a to-do list and tracks its own progress.

📁 Pillar 2: Filesystem — a real Swiss Army knife for AI

A complete set of built-in file operations: SKILL.md and tools like creating folders, categorizing and saving materials, and modifying multiple files.

🐣 Pillar 3: Sub-agents — solving the "not enough brainpower" problem

When a task gets too complex, the main agent can call a task tool to dispatch sub-agents dedicated to specific subtasks. Each sub-agent has its own independent context, with no interference between them.

🧠 Pillar 4: Context Management — a notepad for AI

Automatic summarization plus automatic offloading of large outputs to files solves the "amnesia" problem in long conversations.

🎭 Relationship with Claude Code

DeepAgents explicitly says it is inspired by Claude Code, with key differences:

Claude Code only works with Claude models; DeepAgents supports any LLM
Claude Code is closed-source; DeepAgents is MIT open source
DeepAgents is LangGraph-native, integrating with LangSmith, persistence, and checkpointing

🔧 Out of the Box

One line of code creates an agent, with all the complex features (planning, file operations, sub-agents, context management) enabled out of the box.

📚 Key References

1. LangChain. (2025). DeepAgents. https://github.com/langchain-ai/deepagents 2. ByteIota. (2026). DeepAgents Tutorial. https://byteiota.com/deepagents-tutorial 3. Anthropic. (2025). Claude Code. https://www.anthropic.com/claude-code