What DeerFlow Is

DeerFlow, which stands for Deep Exploration and Efficient Research Flow, is an open-source super agent harness built to orchestrate sub-agents, memory, and sandboxes for complex, multi-step tasks [1]. The framework is designed to handle work that spans minutes to hours, positioning it for long-horizon workflows that exceed the scope of single-turn or short-session agents. ByteDance describes it as capable of doing “almost anything” through extensible skills layered on top of its core orchestration infrastructure [1].

Architecture and Core Components

DeerFlow’s architecture centers on several distinct building blocks that work together to manage task complexity. Sandboxes provide isolated execution environments for code. Memory systems allow agents to retain and reference state across steps. Tools and skills extend what agents can do, while sub-agents handle delegated portions of a larger task. A message gateway coordinates communication across these components [1].

The framework’s skill system is described as extensible, meaning operators can add capabilities beyond the defaults included in the distribution. This design allows teams to tailor DeerFlow to domain-specific workflows without modifying the core orchestration layer.

Version 2.0 Changes

DeerFlow 2.0 is a complete ground-up rewrite. According to the project documentation, it shares no code with version 1 [1]. The original v1 codebase was a Deep Research framework and remains available on the 1.x branch, where contributions are still accepted. Active development, however, has moved entirely to 2.0 [1].

For teams currently running v1, there is no migration path in the conventional sense. Because the two versions share no code, moving to 2.0 effectively means adopting a new system rather than upgrading an existing one. The 1.x branch will continue to receive community contributions, giving operators a window to evaluate 2.0 before committing to a transition.

Supported Models and Integrations

ByteDance recommends three models for running DeerFlow: Doubao-Seed-2.0-Code, DeepSeek v3.2, and Kimi 2.5 [1]. These are listed as the preferred options, though the framework’s setup wizard allows operators to choose from multiple LLM providers during configuration.

On the tooling side, DeerFlow has integrated InfoQuest, an intelligent search and crawling toolset independently developed by BytePlus [1]. InfoQuest supports free online use and is described as handling both search and web crawling functions within the agent’s workflow. This integration is positioned as a first-party option for operators who need retrieval capabilities without configuring a third-party search provider.

Deployment and Setup

DeerFlow supports local deployment and includes Docker support as the preferred execution path [1]. The setup process begins by cloning the repository and running a setup wizard from the project root using make setup. The wizard walks operators through selecting an LLM provider, configuring optional web search, and setting execution and safety preferences including sandbox mode, bash access, and file-write tool permissions. The wizard generates a config.yaml file and writes API keys to a .env file, with the full process estimated at roughly two minutes [1].

For teams using coding agents such as Claude Code, Codex, Cursor, or Windsurf, ByteDance provides a single-prompt bootstrap approach. Operators can pass a one-sentence instruction to the coding agent pointing it to the project’s Install.md file. The agent is designed to clone the repository if needed, select Docker when available, and return the exact next command along with any missing configuration the operator still needs to supply [1].

Who DeerFlow Targets

DeerFlow is aimed at developers and engineering teams building or operating complex, long-running agentic workflows. The framework’s architecture, with its sub-agent orchestration, sandboxed execution, and extensible skill layer, addresses use cases where single-agent or short-context approaches are insufficient. The single-prompt bootstrap and interactive setup wizard lower the barrier for individual developers to get a local instance running quickly, while the Docker support and configurable safety settings address the needs of teams deploying in more controlled environments [1].

FAQ

Q. Can teams continue using DeerFlow v1 after the 2.0 release? Yes. The v1 codebase remains available on the 1.x branch and continues to accept community contributions [1]. ByteDance has not announced a deprecation timeline for the 1.x branch.

Q. Is there a migration path from v1 to v2? No direct migration path exists. DeerFlow 2.0 shares no code with v1, so moving to 2.0 requires adopting the new system from scratch rather than upgrading an existing deployment [1].

Q. Which LLM providers does the setup wizard support? The setup wizard allows operators to choose an LLM provider during configuration, but the sources identify only Doubao-Seed-2.0-Code, DeepSeek v3.2, and Kimi 2.5 as the specifically recommended models [1]. The full provider list is not detailed in available documentation.

Q. Is InfoQuest the only search option available? InfoQuest is the newly integrated first-party option from BytePlus, and the setup wizard includes optional web search configuration [1]. Whether additional third-party search providers are supported is not specified in the available documentation.

Q. Does DeerFlow require Docker to run locally? Docker is the preferred option when available, but the setup wizard guides operators through execution preferences, suggesting non-Docker local deployment is also supported [1].

Key takeaways

  • DeerFlow 2.0 is a complete rewrite with no shared code from the original v1 Deep Research framework; v1 remains on the 1.x branch for continued community contributions [1].
  • The framework orchestrates sub-agents, sandboxed code execution, memory, tools, and extensible skills to handle long-horizon tasks spanning minutes to hours [1].
  • Recommended models are Doubao-Seed-2.0-Code, DeepSeek v3.2, and Kimi 2.5, with BytePlus InfoQuest integrated for search and crawling [1].
  • Local deployment uses a make setup wizard covering LLM provider selection, sandbox mode, and safety preferences, with Docker as the preferred runtime [1].
  • A single-prompt bootstrap approach supports coding agents such as Claude Code, Codex, Cursor, and Windsurf for hands-off setup [1].