Investment Rating - The report does not explicitly state an investment rating for the industry Core Insights - AgentScope is an open-source framework launched by Tongyi Lab, focusing on cutting-edge exploration of agent-related technologies and functionalities, providing production-level solutions for development, training, deployment, and management [3][103] - The framework is built around the core concept of "Agentic," offering four main functionalities: model capability integration, multi-agent orchestration, intelligent context management, and tool management [7][10][19] - AgentScope-Runtime enables Agent-as-a-Service capabilities, allowing agents to be packaged as independently callable API services, facilitating flexible deployment, cost reduction, and rapid iteration [30][31] - The framework supports various deployment protocols, including A2A and Response API, and offers lifecycle management through Docker and Kubernetes [30][32] Summary by Sections AgentScope Features - AgentScope provides a comprehensive set of features including multi-modal models, context management, ReAct paradigm, and multi-agent orchestration [4][10][19] - It supports local deployment and integrates capabilities across text, visual, audio, and multi-modal inputs [14][19] AgentScope-Runtime - The runtime environment supports various deployment methods and includes a tool sandbox for code execution, browser control, and file system services [30][31] - It offers project management capabilities and supports full-link data tracking, providing runtime data statistics [41][63] Evaluation and Observation - AgentScope includes functionalities for agent evaluation, statistical analysis, and visualization of results, enabling quick development of agent templates [33][49][71] - It supports real-time data tracking and project management, enhancing transparency in development [46][63][95] Optimization and Training - AgentScope leverages Trinity-RFT to provide reinforcement learning training capabilities, supporting rapid iteration and optimization of agent applications [99] - It fully supports advanced reinforcement learning algorithms such as SFT, GRPO, GSPO, and PPO, enhancing training efficiency through intelligent sample selection [100]