DeerFlow 2.0: Turning AI Agents into a Super‑Charged, Plug‑and‑Play Harness

ByteDance has open‑sourced DeerFlow 2.0, a project that quickly rose to the top of GitHub Trending and has amassed more than 60.4 k stars. The platform is described as a “Long‑horizon SuperAgent Harness,” a term that has become popular in the AI community.

DeerFlow (Deep Exploration and Efficient Research Flow) is positioned as a turnkey system that makes AI agents genuinely productive. It bundles sub‑agents, memory, and a sandbox environment together and augments them with extensible skills . This combination enables an agent to perform virtually any task, challenging existing solutions such as OpenClaw.

Each task runs inside an isolated Docker container that provides a full file system. Agents can read and write files, execute bash commands, view images, and even generate videos, effectively giving them a personal computer‑like environment.

Skills system : The core of DeerFlow is its skill mechanism. A skill is a Markdown file that defines a specific workflow—research, report generation, PPT creation, web‑page rendering, etc. Users can add custom skills simply by dropping new Markdown files into the skill directory. Skills are loaded on demand, which conserves token usage for large‑language‑model contexts.

Sub‑Agents parallel execution : For complex objectives, the main agent decomposes the problem into multiple sub‑agents that run concurrently. After all sub‑agents finish, their results are merged into a final answer. For example, an “AI industry trend report” can be split into sub‑tasks covering large models, multimodal data, open‑source ecosystems, and commercialization, dramatically speeding up the process.

Sandbox isolation : The sandbox is a true Docker container with a fixed directory structure, ensuring that different sessions do not interfere with each other. DeerFlow supports three deployment modes—local, Docker, and Kubernetes—making it ready for production use.

Context engineering : Within a single session, DeerFlow actively manages context by summarizing completed sub‑tasks, persisting intermediate results to the file system, and compressing less‑important information. This prevents the model’s context window from being exhausted during long, multi‑step workflows.

Long‑term memory : Across sessions, DeerFlow gradually builds a persistent memory store that captures user preferences, knowledge background, and habitual workflows. The memory is stored locally, giving users full control over their data.

Getting Started

To try DeerFlow, clone the repository and follow the installation instructions in the raw Install.md file.

If you haven’t cloned DeerFlow yet, clone it first, then run the commands in https://raw.githubusercontent.com/bytedance/deer-flow/main/Install.md to set up the local development environment.

https://github.com/bytedance/deer-flow/blob/main/README_zh.md