DeerFlow

• 1. overview (636 chars)

  DeerFlow is a super Agent framework open sourced by ByteDance. Its full name is Deep Exploration and Efficient Research Flow. The framework was first open sourced in May 2025 and quickly gained traction on GitHub, with more than 39,000 stars as of March 2026. The positioning of DeerFlow version 2.0 has been upgraded from a mere research tool to a true Super Agent Harness, capable of completing complex tasks that require dozens or even hundreds of steps.   Different from traditional Agent frameworks, DeerFlow provides developers with a set of production-level artificial intelligence agent solutions through innovative designs such as Skills system, dynamic Sub-Agent orchestration, isolated sandbox execution environment and long-term memory.

• 2. Product background (802 chars)

  DeerFlow was developed and launched by ByteDance and is an important open source contribution of China's leading technology company in the field of AI Agent. The project was born out of an insight into the limitations of traditional AI assistants: when AI models such as ChatGPT are capable of generating human-level code and text, the bigger challenge is how to get AI to complete complex tasks that require multi-step, cross-tool, and long execution times. Traditional AI assistants are typically single-interaction models that cannot effectively handle workflows that require in-depth research, coding, file manipulation, and multiple iterations.   Based on its deep accumulation in the field of artificial intelligence, ByteDance officially open sourced the DeerFlow project on GitHub on May 9, 2025. In March 2026, DeerFlow version 2.0 was released, and its positioning was upgraded from a deep research framework to a super agent framework.

• 3. Product functions and usage experience (1938 chars)

  DeerFlow's core functionality can be understood along five main dimensions.   The first dimension is the SKILLS & TOOLS system. This is one of DeerFlow’s most unique innovations. Traditional frameworks usually require defining skills through code, while DeerFlow uses Markdown files to define skills, which greatly reduces the expansion threshold. Built-in skills include: research (in-depth research), report (generating structured reports), slide (creating slides), webpage (generating web pages), image (generating pictures), and video (generating videos). The core toolset includes web_search (web search), web_fetch (web page scraping), file_operations (file reading and writing), and bash (Shell command execution). The system also supports MCP Servers integration and Python function extensions.   The second dimension is the Sub-Agent architecture. When a user submits a complex task, the Lead Agent will first decompose the task and then dynamically generate multiple Sub-Agents for parallel execution. Each Sub-Agent has an independent context environment, which avoids the context pollution problem of traditional single large model solutions. After execution is complete, the Lead Agent is responsible for integrating all results and generating the final output.   The third dimension is sandbox and file system. DeerFlow provides an isolated execution environment for each task, which is an important security feature. The working directory of a task contains three main subdirectories: uploads (user uploaded files), workspace (workspace, readable and writable), and outputs (generated output files). The system supports three execution modes: Local (local direct execution), Docker (container isolation) and Kubernetes (suitable for enterprise-level cluster deployment).   The fourth dimension is Context Engineering. The system adopts an isolated Sub-Agent Contexts design. Each sub-agent has an independent context window and dedicated tool set to prevent the main Agent's context from being contaminated. Long-term memory functionality supports persistence across sessions.   The fifth dimension is multiple output formats. DeerFlow supports more than 7 output formats, including reports, slideshows, web pages, images, videos, etc.

• 4. Pricing and business model (812 chars)

  DeerFlow is an open source, free project licensed under the MIT license. Users can clone the repository and deploy it on GitHub for free. There is no fee for the project itself.   As an open source project, DeerFlow's business model is mainly realized through the following methods: first, technical influence, companies establish technology brands and developer communities through open source projects; second, enterprise service support, ByteDance may provide companies with paid technical support, customized development and training services; third, talent attraction, open source projects are an important channel to attract outstanding technical talents.   From a cost perspective, the costs that users need to bear when deploying DeerFlow mainly include: cloud servers or physical machines (used to run Docker containers), model API calling fees (if paid models such as OpenAI and Anthropic are used), and development and maintenance labor costs.

• 5. User reviews (916 chars)

  Judging from community feedback, DeerFlow has received positive user reviews.   The main positive comments include: the framework design is advanced, and the concepts of Skills and Sub-Agent are clear and easy to understand; deployment is relatively simple, and the Docker containerization solution lowers the deployment threshold; it is powerful and can handle complex practical tasks; the documentation is complete, and the README and technology blog provide detailed instructions.   The main negative feedback includes: it requires a certain technical background to deploy and configure correctly; it has certain requirements for hardware resources and the running cost cannot be ignored; some functions need to be configured to achieve maximum effectiveness; there are currently relatively few Chinese documents.   Typical usage scenarios include: in-depth research analysis and content generation, data processing workflow, enterprise system integration (supporting IM channel access such as Telegram, Slack, Feishu, etc.), automated process orchestration, etc.

• 6. Industry evaluation (851 chars)

  From an industry perspective, the release of DeerFlow marks the open source competitiveness of Chinese technology companies in the AI ​​Agent field. Compared with international mainstream frameworks, DeerFlow has unique innovations in features such as dynamic generation of Sub-Agent, Markdown skill definition, and isolation sandbox.   DeerFlow is rated highly by technology bloggers and the developer community. The technical analysis article published by iTech in the blog park details the architectural design and implementation details of DeerFlow. Xiago Programming’s technical article provides an in-depth analysis of DeerFlow’s enterprise-level deployment solution.   From the competitive product landscape, DeerFlow mainly competes with frameworks such as LangChain, AutoGPT, and CrewAI. The advantage of DeerFlow lies in dynamic Sub-Agent, Markdown skill definition and built-in sandbox file system, which is more suitable for complex enterprise-level application scenarios.

• 7. Controversies and Risks (602 chars)

  As a product of ByteDance, DeerFlow may face the following controversies and risks:   Geopolitical sensitivity: As an open source project of a Chinese company, some overseas users may be wary of deployment due to data security concerns.   Security challenges: Although the sandbox provides an isolation environment, capabilities such as shell command execution still require careful configuration of whitelists, and enterprise users need to evaluate security risks.   Long-term maintainability: The long-term maintenance of open source projects depends on the company's investment.   Learning curve: Although the concept is advanced, proper deployment and use still requires a certain technical background.

• 8. Applicable groups and usage suggestions (762 chars)

  DeerFlow is suitable for the following user groups: enterprise developers whose technical teams need to build AI Agent systems, individual developers who have needs for AI automation, researchers who conduct AI research and experiments, and content creators who need in-depth research capabilities.   For different usage scenarios, the recommendations are as follows: Docker Gateway Mode can be used directly for personal learning and experimentation; it is recommended to use Standard Mode and configure a Kubernetes cluster for enterprise production environments; for scenarios that require IM channel integration, Telegram or Feishu access can be configured.   In terms of alternatives, if DeerFlow does not meet your needs, you can consider: LangChain (more mature, richer documentation), LangGraph (lower level, suitable for deep customization), CrewAI (simpler Agent orchestration).

• 9. Summarize (480 chars)

  DeerFlow is a production-level Super Agent framework open sourced by Bytedance. It provides new ideas for building AI Agents through innovative designs such as dynamic Sub-Agent orchestration, Markdown skill definition, isolated sandbox execution environment, and long-term memory. For businesses and developers who need to build complex AI automation systems, DeerFlow is an option worth considering. With the release of version 2.0, the capabilities of the framework have been significantly enhanced, and future development deserves continued attention.

No in-depth report yet

• 

  CosmosCris48

  —

  用了两周 DeerFlow，不得不说 Slabs 编排确实香，但文档太少了，很多功能得自己踩坑。

---

• 

  枫叶714

  —

  50K Star 不是吹的，2.0 版本比 1.0 强太多了！自动拆解任务并行执行，省了我不少事儿。

---

• 

  BenjaminBrownQ

  —

  实测 Sub-Agent 能力不错，但模型幻觉问题还是存在，需要加个校验流程。

---

• 

  Russell.Roberts369

  —

  部署是真的麻烦，文档写得不清不楚，建议官方出点视频教程。

---

• 

  Jennifer_BrooksZ

  —

  Docker 沙箱隔离这个设计很棒，安全有保障。

---

• 

  RRoss_2022418

  —

  比 LangChain 好上手多了，技能系统很直观。

---

• 

  RWalker_2022

  —

  mark 一下，等成熟了再用。

---

• 

  Harold_Nguyen_725

  —

  企业级应用还是算了吧，个人开发者玩一玩还行。

---

• 

  Nicholas_FloresII

  —

  yyds！终于有个能打的国产 Agent 框架了。

---

• 

  JClark520

  —

  GitHub Trending 第一！字节这次真的给力。

---

• 

  MBell_77

  —

  刚部署成功，界面挺干净的，希望能满足我的研究需求。

---

• 

  AAllen

  —

  长期记忆功能好用，跨会话能记住我之前的研究内容。

---

• 

  SandyFreier

  —

  skill 定义用 Markdown 确实方便，非程序也能改。

---

• 

  AlexaStewart

  —

  研究了一下午，发现这玩意儿还挺复杂，新手劝退。

---

• 

  TIalv

  —

  比 AutoGPT 稳定多了，期待后续更新。

---

• 

  掠影_12

  —

  多模型接入好评，能用 Claude 也能用 GPT。

---

• 

  MatthewNelsonK

  —

  生成的报告格式挺好看，直接能用。

---

• 

  David_WrightZ_196

  —

  官方 Demo 运行成功，但自定义 skill 一直报错，麻了。

---

• 

  NicoleJames520

  —

  从研究工具升级到 Super Agent，这波升级可以的。

---

• 

  Natalie.Kim_2020

  —

  并行执行 sub-agent 效率很高，复杂任务可以放心交给它。

No reviews yet