**Abstract:** AgentScope is a developer-centric multi-agent platform designed to address the challenges of coordinating agents' cooperation and managing the erratic performance of Large Language Models (LLMs). It features a message exchange mechanism, built-in agents, service functions, user-friendly interfaces, zero-code programming, and automatic prompt tuning. The platform supports robust and flexible multi-agent applications with fault tolerance mechanisms, multi-modal data management, and distributed deployment capabilities. AgentScope aims to lower the barriers to development and deployment, making it easier for developers to build intelligent agent applications. **Introduction:** Multi-agent systems require collaborative efforts from multiple agents, which poses significant challenges in development. AgentScope addresses these challenges by providing a user-friendly platform with a message exchange mechanism, built-in utilities, and automatic prompt tuning. It also includes robust fault tolerance mechanisms, support for multi-modal data, and tools for efficient distributed deployment. **Core Concepts:** - **Message:** Messages are the carriers of information exchange in multi-agent conversations. - **Agent:** Agents are the primary actors in multi-agent applications, handling tasks and interactions. - **Service:** Service functions are functional APIs that return formatted outputs. - **Workflow:** Workflows define the sequence of agent executions and message exchanges. **Architecture:** AgentScope's architecture consists of three layers: Utility Layer, Manager and Wrapper Layer, and Agent Layer. Each layer supports different functionalities, ensuring high availability and resilience. **High Usability:** - **Syntactic Sugar:** Abstracts complex message exchanges with pipelines and message hubs. - **Resource-Rich Environment:** Provides built-in resources, pre-built agents, and demonstration interfaces. - **Drag-and-Drop Programming:** Offers a zero-code programming workstation for easy application development. - **Automatic Prompt Tuning:** Automates prompt generation and tuning for LLMs. **Fault-Tolerant Mechanisms:** - **Error Classification:** Categorizes errors into accessibility, rule-resolvable, model-resolvable, and unresolvable. - **Handling Strategies:** Includes auto-retry, rule-based correction, customizable fault handlers, and agent-level fault handling. **Multi-Modal Applications:** - **Management:** Supports generation, storage, and transmission of multi-modal data. - **Interaction Modes:** Facilitates multi-modal data interaction through terminal and web UI. **Tool Usage:** - **Function Preparation:** Preprocesses service functions for LLMs. - **Instruction Preparation:** Prepares tool instructions and calling formats. - **Iterative Reasoning:** LLMs generate strategic reasoning and decisions. - **Iterative Acting:** Parses and executes LLM responses. **Retrieval-Augmented Generation:** - **RAG Methodology:** Enhances LLMs with pre-processing steps to access customized knowledge domains. **Conclusion:** AgentScope is a comprehensive platform that simplifies the development and deployment of multi-agent applications, leveraging the capabilities of LLMs. It offers robust features and tools to handle complex interactions and**Abstract:** AgentScope is a developer-centric multi-agent platform designed to address the challenges of coordinating agents' cooperation and managing the erratic performance of Large Language Models (LLMs). It features a message exchange mechanism, built-in agents, service functions, user-friendly interfaces, zero-code programming, and automatic prompt tuning. The platform supports robust and flexible multi-agent applications with fault tolerance mechanisms, multi-modal data management, and distributed deployment capabilities. AgentScope aims to lower the barriers to development and deployment, making it easier for developers to build intelligent agent applications. **Introduction:** Multi-agent systems require collaborative efforts from multiple agents, which poses significant challenges in development. AgentScope addresses these challenges by providing a user-friendly platform with a message exchange mechanism, built-in utilities, and automatic prompt tuning. It also includes robust fault tolerance mechanisms, support for multi-modal data, and tools for efficient distributed deployment. **Core Concepts:** - **Message:** Messages are the carriers of information exchange in multi-agent conversations. - **Agent:** Agents are the primary actors in multi-agent applications, handling tasks and interactions. - **Service:** Service functions are functional APIs that return formatted outputs. - **Workflow:** Workflows define the sequence of agent executions and message exchanges. **Architecture:** AgentScope's architecture consists of three layers: Utility Layer, Manager and Wrapper Layer, and Agent Layer. Each layer supports different functionalities, ensuring high availability and resilience. **High Usability:** - **Syntactic Sugar:** Abstracts complex message exchanges with pipelines and message hubs. - **Resource-Rich Environment:** Provides built-in resources, pre-built agents, and demonstration interfaces. - **Drag-and-Drop Programming:** Offers a zero-code programming workstation for easy application development. - **Automatic Prompt Tuning:** Automates prompt generation and tuning for LLMs. **Fault-Tolerant Mechanisms:** - **Error Classification:** Categorizes errors into accessibility, rule-resolvable, model-resolvable, and unresolvable. - **Handling Strategies:** Includes auto-retry, rule-based correction, customizable fault handlers, and agent-level fault handling. **Multi-Modal Applications:** - **Management:** Supports generation, storage, and transmission of multi-modal data. - **Interaction Modes:** Facilitates multi-modal data interaction through terminal and web UI. **Tool Usage:** - **Function Preparation:** Preprocesses service functions for LLMs. - **Instruction Preparation:** Prepares tool instructions and calling formats. - **Iterative Reasoning:** LLMs generate strategic reasoning and decisions. - **Iterative Acting:** Parses and executes LLM responses. **Retrieval-Augmented Generation:** - **RAG Methodology:** Enhances LLMs with pre-processing steps to access customized knowledge domains. **Conclusion:** AgentScope is a comprehensive platform that simplifies the development and deployment of multi-agent applications, leveraging the capabilities of LLMs. It offers robust features and tools to handle complex interactions and