⌘K
Change language Switch ThemeSign In
Narrow Mode
字节版龙虾架构火爆 GitHub！开源获 35k+ Star，内置 Skill 全家桶，原生适配飞书
量 量子位 @闻乐
One Sentence Summary
ByteDance's open-source Deer-Flow2 is a modular multi-agent management framework based on LangGraph. With its innovative middleware architecture, isolated sandbox environment, and native IM integration, it has rapidly gained 35k+ stars on GitHub.
Summary
This article provides an in-depth analysis of Deer-Flow2, the latest open-source super-agent management framework from ByteDance. Version 2.0 features a complete architectural overhaul, evolving from a fixed multi-node model to a flexible structure of "Single Master Agent + 11-Layer Middleware Chain + Dynamic Sub-Agents." Key highlights include: a built-in "Skill Suite" with various search and crawler tools, a Docker-based AIO Sandbox for secure execution, context compression and scheduling mechanisms for long-running tasks, and native support for IM channels like Lark and Slack. The article not only analyzes its technical principles but also provides detailed Docker and local deployment guides, demonstrating its practical capabilities in scenarios such as web generation and creative design.
Main Points
* 1. Architecture evolves from fixed nodes to "middleware-driven."Version 2.0 consolidates core capabilities through an 11-layer middleware chain, supporting dynamic scheduling of sub-agents. This allows the system to remain lightweight while offering high extensibility, enabling the addition of new skills without modifying the underlying code. * 2. Built a pluggable, progressive skill system.Includes over a dozen built-in skills such as deep research and data analysis, and supports custom tools via MCP or Python interfaces. The system dynamically loads skills based on requirements to optimize token consumption. * 3. Introduced an isolated sandbox to ensure secure code execution.Utilizes ByteDance's open-source AIO Sandbox to provide an independent, isolated environment for each task, supporting full file system operations and Bash execution permissions, ensuring consistency between development and deployment environments. * 4. Systematically addresses context bottlenecks in long-running tasks.Effectively handles window limitations in complex tasks through mechanisms such as structured task decomposition, parallel sub-agent scheduling, automatic context summarization/compression, and external storage.
Metadata
AI Score
87
Website qbitai.com
Published At Today
Length 2249 words (about 9 min)
Sign in to use highlight and note-taking features for a better reading experience. Sign in now
> 闻乐 发自 凹非寺 > > > 量子位 | 公众号 QbitAI
原来字节也有龙虾—— Deer-Flow2超级智能体管理框架。
开源发布后迅速登上了GitHub Trending榜首，已经收获了35.3k Star。
Deer-Flow2采用模块化多智能体架构，这些智能体通过LangGraph实现协同合作。
主打开箱即用，内置了Tavily、Brave Search、DuckDuckGo等多种搜索引擎，还集成了Jina等爬虫工具，基本把信息收集的十八般兵器都给配齐了。
当然，扩展性也没落下，自定义API或模型随意接。
核心能力上，多智能体协同、沙箱安全执行、一键部署全都有，Docker快速部署和本地开发任你挑，主流大模型统统兼容。
不过最贴心的还得是IM渠道支持——
原生适配飞书、Telegram、Slack，没有公网IP也能跑。
DeerFlow在迭代过程中完成了一次彻底的架构升级。
1.0版本采用固定5节点多智能体架构，能力边界相对明确，主要聚焦在深度研究场景。
而2.0版本则对整体结构进行了全面重构，从底层骨架到上层能力都实现了飞跃。
新版本采用单一主智能体+11 层中间件链+动态子智能体的全新架构，将核心能力收敛到工具集与中间件链中，让整个系统更轻量、更灵活、更易扩展。
相比1.0需要调整整体结构才能新增能力，2.0只需添加新技能就能完成拓展，无需改动底层框架。
原本作为核心的深度研究，也从唯一主打能力转变为框架内置的一项基础能力。
在框架层面，DeerFlow 2.0已经整合子智能体调度、长期记忆、隔离沙箱执行环境、可扩展技能与工具等关键模块，形成了一套完整、成熟的智能体运行能力体系。
为了让智能体快速适配不同场景，DeerFlow 2.0搭建了一套可插拔的技能体系。
出厂自带深度研究、数据分析、图表生成、音视频创作等十余种常用技能，系统会根据任务需求渐进式加载控制token消耗，这样就避免了上下文被过度占用而导致的效率下降。
如果内置能力无法满足需求，用户还可以自行封装专属技能。
配合官方提供的skill-creator工具，几分钟就能为智能体扩展新能力。
同时系统提供MCP与Python接口，支持自定义工具的深度集成，甚至可接入Claude Code，让用户在终端就能完成工具的下发、查看与管理操作
DeerFlow 2.0还配备了独立隔离沙箱。每个任务都在专属沙箱中运行，拥有完整文件系统与Bash执行权限，支持文件读写、脚本运行、命令操作等。
系统提供本地、Docker、Kubernetes三种运行模式。
其中Docker模式采用字节开源的AIO Sandbox，隔离级别更高、运行更稳定。
同时自动完成虚拟路径与物理路径的映射，确保开发环境与部署环境保持一致。
面对复杂长时任务，DeerFlow 2.0通过调度机制与上下文工程双管齐下。
主智能体会先对任务进行结构化拆解，再按需调度最多3个子智能体并行执行子智能体可选用通用能力或命令行专家型。
每个子智能体都拥有独立上下文，互不干扰、互不污染
在此基础上，框架还通过多层中间件链、上下文自动摘要压缩、外部文件存储、子任务限流等设计，系统性解决长时任务中上下文窗口不足的问题。
说了这么多，接下来检验一下DeerFlow 2.0的能力如何。
一键产出完整、可交付的足球联赛官网页面，从设计到代码全流程自动化。
一句指令就能把复杂概念变成孩子也能看懂的哆啦A梦漫画！
一句话生成液态玻璃天气界面，鼠标悬停还能3D形变。
DeerFlow提供了Docker和本地这两种主要的部署方式。
Docker部署是最简单快捷的方式，只需几个命令，就能在本地启动完整的DeerFlow服务。
首先克隆仓库：
> git clone https://github.com/bytedance/deer-flow.git
等待仓库下载完成后，进入项目根目录：
> cd deer-flow
生成本地配置，输入：
> make config
系统会自动生成config.yaml配置文件和.env文件（如果没有make命令，Windows可安装MinGW）。
然后找到项目目录下的Config.yaml文件，填入模型相关配置。
设置完成后，输入：
> make docker-init
自动拉取字节开源的AIO Sandbox沙箱镜像，首次拉取可能需要几分钟。
镜像拉取完成后，启动服务，输入：
> docker-start
服务启动后，访问http://localhost:2026即可进入Web界面。
如果需要进行深度定制或二次开发，可以选择本地部署方式。
本地部署需要满足一定的前置条件，包括Python 3.12+、Node.js 22+、pnpm、uv包管理器以及nginx。
满足前置环境后检查依赖，打开终端进入deer-flow根目录，输入：
> make check
系统会自动校验上述依赖是否齐全、缺少的会提示补充。
这时候你可以输入：
> make install
系统会自动安装python和node相关依赖包。
接下来可以输入make setup-sandbox（按需）预拉取沙箱镜像，避免后续首次使用时等待。
然后启动服务：
> make dev
这种方式适合需要修改源码、调试功能或贡献代码的开发者。
DeerFlow原生支持从即时通讯应用接收任务，目前支持Telegram、Slack和飞书/Lark三个渠道，且都不需要公网IP。
!Image 34△config.yaml文件channels相关配置
配置完成后，就可以直接在聊天窗口中与DeerFlow交互。
DeerFlow的两位核心开发者是来自北京大学的Tao He和来自南京大学的Henry Li。
项目地址：https://github.com/bytedance/deer-flow
官方网站：https://deerflow.tech
参考链接：https://x.com/Gorden_Sun/status/2035698488034628003
— 完 —
量 量子位 @闻乐
One Sentence Summary
ByteDance's open-source Deer-Flow2 is a modular multi-agent management framework based on LangGraph. With its innovative middleware architecture, isolated sandbox environment, and native IM integration, it has rapidly gained 35k+ stars on GitHub.
Summary
This article provides an in-depth analysis of Deer-Flow2, the latest open-source super-agent management framework from ByteDance. Version 2.0 features a complete architectural overhaul, evolving from a fixed multi-node model to a flexible structure of "Single Master Agent + 11-Layer Middleware Chain + Dynamic Sub-Agents." Key highlights include: a built-in "Skill Suite" with various search and crawler tools, a Docker-based AIO Sandbox for secure execution, context compression and scheduling mechanisms for long-running tasks, and native support for IM channels like Lark and Slack. The article not only analyzes its technical principles but also provides detailed Docker and local deployment guides, demonstrating its practical capabilities in scenarios such as web generation and creative design.
Main Points
* 1. Architecture evolves from fixed nodes to "middleware-driven."
Version 2.0 consolidates core capabilities through an 11-layer middleware chain, supporting dynamic scheduling of sub-agents. This allows the system to remain lightweight while offering high extensibility, enabling the addition of new skills without modifying the underlying code.
* 2. Built a pluggable, progressive skill system.
Includes over a dozen built-in skills such as deep research and data analysis, and supports custom tools via MCP or Python interfaces. The system dynamically loads skills based on requirements to optimize token consumption.
* 3. Introduced an isolated sandbox to ensure secure code execution.
Utilizes ByteDance's open-source AIO Sandbox to provide an independent, isolated environment for each task, supporting full file system operations and Bash execution permissions, ensuring consistency between development and deployment environments.
* 4. Systematically addresses context bottlenecks in long-running tasks.
Effectively handles window limitations in complex tasks through mechanisms such as structured task decomposition, parallel sub-agent scheduling, automatic context summarization/compression, and external storage.
Key Quotes
* The new version adopts a brand-new architecture consisting of a single master agent, an 11-layer middleware chain, and dynamic sub-agents, consolidating core capabilities into the toolset and middleware chain. * The system progressively loads skills based on task requirements to control token consumption, thereby avoiding efficiency drops caused by excessive context usage. * Each task runs in a dedicated sandbox with full file system and Bash execution permissions, supporting file read/write, script execution, command operations, and more. * Native support for Lark, Telegram, and Slack; it can even run without a public IP address.
AI Score
87
Website qbitai.com
Published At Today
Length 2249 words (about 9 min)
Tags
Deer-Flow2
Multi-Agent Framework
ByteDance
LangGraph
AI Agent
Related Articles
* Yao Shunyu Lectures Face-to-Face with Tang Jie, Yang Zhilin, and Lin Junyang! Four Schema Heroes Debate Heroes at Zhongguancun * What You Don't Know About Claude Code: Architecture, Governance, and Engineering Practices - Tw93 * 128. Manus Co-founder's Final Interview Before Sale: The Fantastical Drift to 2025... * GPT-5.4 Released: OpenAI's First Unified Model, Truly Native * MiniMax M2.1 Emerges as a Leader in Multilingual AI Coding, Achieving SOTA * After Topping Open-Source Rankings with its Programming LLM, the Zhipu GLM Team Faced a 3-Hour Questioning Session * Vol.89 AI Industry 2025 Annual Summary Supplement (V4 "Can't Wait" Edition) --- 70-page PPT Solo * Deconstructing Clawdbot: Local Architecture, Memory Management, Agent Orchestration, and Context Assembly Principles * AI Starts to "Take Action", Alibaba's Qwen Leads the World * How to Design an AI Agent System HomeArticlesPodcastsVideosTweets