DeerFlow 2.0 is an open-source long-horizon AI agent framework released by ByteDance in February 2026. Similar to a film director simultaneously managing cameras, lighting, sound, and actors on a film set, DeerFlow's SuperAgent decomposes heterogeneous tasks such as research, coding, website creation, and presentation generation into sub-agents and performs them concurrently through parallel orchestration. Built on top of LangGraph and LangChain, each sub-agent runs independently within an isolated Docker container, with its own file system, Bash terminal, and browser access. It is a complete rewrite that shares no code with v1.x, and within 24 hours of its release, it topped the GitHub Trending list and currently boasts 71.4k stars, demonstrating one of the highest adoption rates among open-source agent frameworks. Existing agent frameworks each had distinct limitations. LangChain excels in fine-grained control at the chat level but requires building production infrastructure directly, AutoGPT is attractive for experimentation but lacks stability during long-term execution, and CrewAI specializes in role-based agent collaboration but has static role definitions. DeerFlow 2.0 overcomes all three of these limitations simultaneously. The sandbox execution environment completely isolates each task, preventing contamination of the host system, and persistent memory between sessions preserves user preferences and accumulated knowledge for the long term. The Skills Framework supports Markdown-based skill definitions and enables slash commands (e.g., /data-analysis analyze file.csv), and progressively loads skills based on task requirements to optimize token usage. It natively supports the MCP (Model Context Protocol) server and can be integrated with all OpenAI-compatible models with 100k+ context windows, including OpenAI, Anthropic Claude, DeepSeek, Google Gemini, and Qwen. From the perspective of a life science researcher, DeerFlow 2.0 is particularly useful for long-term autonomous research workflows. For example, dozens of recent papers on a specific target protein can be collected in parallel, with each sub-agent independently performing summarization and citation extraction, and the SuperAgent then synthesizes these into a comprehensive report within a single conversation session. Through the coding agent, statistical analysis scripts for the collected data can be automatically generated and immediately executed and validated within the sandbox, and the results can be converted into presentation slides or web dashboards, all within a single framework. It can also be integrated with six messaging platforms, including Telegram, Slack, Feishu (Lark), WeChat, and DingTalk, allowing researchers to directly invoke the agent from their team's channels. Furthermore, it has built-in enterprise-grade observability through LangSmith and Langfuse integration, allowing for the tracking of all LLM calls and tool executions.

ByteDance's open-source long-term task AI agent framework.

Autonomous Research Agent: Multiple sub-agents perform parallel web searches, collect papers, and analyze competitors for a specific research topic, and the SuperAgent automatically generates a comprehensive report including citations. By processing dozens of sources simultaneously, the time required is significantly reduced compared to manual work.