ByteDance has open-sourced DeerFlow 2.0, a super-agent orchestration framework designed for complex, long-horizon tasks that automates multi-step workflows.

It allows primary agents to dynamically dispatch specialized sub-agents for tasks like coding or research, while using isolated sandboxes to ensure safe code execution. This maintains context consistency and forms a 'research-execute-verify' loop for tasks spanning minutes to hours.

Why does DeerFlow 2.0 matter for modern AI engineering and content pipelines?

It bridges foundation models and real-world engineering by combining deep exploration with efficient workflows, establishing a new benchmark for reliable multi-agent collaboration and significantly boosting AI autonomy in complex R&D scenarios.

What security risks should developers manage, and what are the next expected steps for the framework?

Security risks like sandbox escapes must be managed. Future updates will focus on integrating specialized tools and optimizing inter-agent communication efficiency to reduce latency.

字節跳動開源 DeerFlow 2.0：基於子智能體與沙箱的超長週期超級智能體框架

DeerFlow 2.0 是字節跳動開源的超級智能體編排框架，專為複雜長週期任務設計。通過調度子智能體、持久化記憶、隔離沙箱和可擴展技能模組，框架能夠自動處理需數分鐘至數小時的研究、編碼和內容創作任務。其核心優勢在於深度探索與高效研究的流程結合，支援程式碼執行、網頁爬取及多輪交互，顯著提升 AI 在真實工程場景中的自主性與可靠性，適用於研發與內容生產等需要深度資訊檢索和複雜邏輯推理的場景。