DeerFlow 2.0 — ByteDance's Long-Running SuperAgent Harness

ByteDance has released a complete rewrite of DeerFlow 2.0. It shares no framework or code with v1's Deep Research framework, and built on top of LangGraph and LangChain, it bundles sub-agents to decompose and parallelize long-running tasks such as research, coding, and content generation.

The built-in Markdown-based Agent Skill system covers research, reports, slides, web pages, and image/video generation. Custom skills can be added, swapped, or combined, and it's designed with progressive loading that only loads skills when needed to reduce token usage.

Task execution happens in an independent sandbox for each task.

• Filesystem: uploads/workspace/outputs
• Features: shell execution, image viewing
• AioSandboxProvider: runs in an isolated Docker container
• LocalSandboxProvider: host bash disabled by default

Operational features have also been strengthened.

• Long-Term Memory persists user profiles, preferences, and tech stacks locally across sessions
• Automatic skipping of duplicate entries is supported
• Per-sub-agent isolated context, completed task summaries, intermediate result filesystem offloading, and Strict Tool-Call Recovery prevent exceeding the context window

Integration coverage is also broad.

• Support for 6 IM channels: Telegram, Slack, Feishu/Lark, WeChat, WeCom, DingTalk
• MCP server extensions and OAuth token flow support
• The claude-to-deerflow skill enables direct integration from the Claude Code terminal
• The make setup wizard completes initial setup in about 2 minutes
• Support for various LLM providers including OpenAI, OpenRouter, vLLM, Codex CLI, Claude Code OAuth
• Simultaneous tracing with LangSmith and Langfuse
• An Embedded Python Client is provided
• MIT license