字节版龙虾架构火爆GitHub！开源获35k+ Star，内置Skill全家桶，原生适配飞书

Core Insights - The article discusses the launch of Deer-Flow2, a modular multi-agent management framework developed by ByteDance, which has quickly gained popularity on GitHub with 35.3k stars [2]. Group 1: Framework Features - Deer-Flow2 utilizes a modular multi-agent architecture, enabling agents to collaborate through LangGraph, and comes pre-equipped with various search engines and crawling tools [4]. - The framework supports extensibility, allowing users to customize APIs or models easily [5]. - Key capabilities include multi-agent collaboration, sandbox security execution, and one-click deployment, compatible with mainstream large models [6]. - The framework supports native integration with communication channels like Feishu, Telegram, and Slack, allowing operation without a public IP [7]. Group 2: Version Improvements - Version 1.0 featured a fixed 5-node multi-agent architecture focused on deep research scenarios [10]. - Version 2.0 has undergone a complete structural overhaul, adopting a new architecture with a single main agent, 11 middleware layers, and dynamic sub-agents, making the system lighter, more flexible, and easier to expand [11]. - The core capability of deep research has transitioned to a foundational ability within the framework, which now includes key modules like sub-agent scheduling and long-term memory [13]. Group 3: Skill System - Deer-Flow 2.0 features a pluggable skill system, pre-loaded with over ten common skills such as deep research, data analysis, and multimedia creation, which can be incrementally loaded based on task requirements [15]. - Users can create custom skills using the provided skill-creator tool, allowing for rapid expansion of agent capabilities [18]. Group 4: Execution Environment - The framework includes an independent isolated sandbox for each task, providing a complete file system and Bash execution permissions, supporting file read/write and script execution [20]. - It offers three operational modes: local, Docker, and Kubernetes, with Docker mode providing higher isolation and stability [21][22]. Group 5: Task Management - Deer-Flow 2.0 employs a scheduling mechanism and context engineering to manage complex long-duration tasks, allowing the main agent to structure tasks and dispatch up to three sub-agents for parallel execution [25]. - Each sub-agent operates in an independent context, preventing interference and pollution, while the framework addresses context window limitations through various design features [27]. Group 6: Deployment Instructions - The article provides detailed instructions for deploying Deer-Flow using Docker or locally, with Docker being the simpler option requiring minimal commands [33][39]. - Local deployment requires specific prerequisites and allows for source code modification and debugging [40][48]. Group 7: Communication Integration - Deer-Flow natively supports task reception from instant messaging applications, specifically Telegram, Slack, and Feishu/Lark, without needing a public IP [49].