Delegate hours of research, coding, and production to a single instruction
Imagine handing off tasks that take hours, from complex research to code writing and content production, with a single instruction and waiting for the results. DeerFlow is an open-source super agent harness that coordinates multiple sub-agents and performs tasks in a sandbox environment. Version 2.0 has been rewritten with a structure completely separate from the existing code to handle long-duration tasks more stably.
Complete rewrite of version 2.0 and recent trends
DeerFlow 2.0 is a ground-up rewrite that shares no code with the 1.x version. The 1.x version is maintained as a 'deep research' framework and accepts contributions, but active development has moved to 2.0. On February 28, 2026, following the release of version 2.0, it reached number one on GitHub Trending, attracting community attention. This version is designed to perform almost any task by orchestrating sub-agents, memory, and sandboxes based on extensible skills.
Task processing methods and integrated tools
The system combines sandboxes, memory, tools, skills, sub-agents, and message gateways to handle various levels of tasks ranging from minutes to hours.
- Multi-agent orchestration A lead agent coordinates sub-agents to decompose and execute complex tasks.
- Sandbox and file system Executes code and manipulates files in a safe, isolated environment.
- Long-term memory and context management Maintains long conversations and tasks through session goals, manual context compression, and long-term memory features.
- Browser control and RAG Supports agentic browser control and private knowledge retrieval via RAGFlow.
Installation and execution methods
Using Docker is recommended, and setting up a local development environment is also possible. If you are using coding agents such as Claude Code, Codex, Cursor, or Windsurf, you can automate everything from cloning to bootstrapping with a one-line instruction. It is strongly recommended to use the ByteDance Volcengine Doubao-Seed-2.0-Code, DeepSeek v3.2, and Kimi 2.5 models. The HTTP connection/read timeout for InfoQuest is set to 30 seconds and operates independently of the crawling timeout.
Security and tracing settings for deployment
Gateway management privileges are equivalent to code execution, so special attention to security is required. Inappropriate deployment can lead to security risks, and external chat message role settings and deployment defaults should be reviewed carefully. It supports various tracing options such as LangSmith, Langfuse, and Monocle to facilitate request tracking and debugging. You can comply with provider requests per minute limits through the request_admission setting for specific models, which is disabled by default.
