The paper introduces AIOS, an innovative operating system designed to manage and optimize the deployment of LLM-based agents. AIOS addresses the challenges of resource management and scheduling in agent systems, which can lead to inefficient or harmful resource allocation. The architecture of AIOS is divided into three layers: the application layer, the kernel layer, and the hardware layer. The application layer provides an SDK for developers to build and integrate agents, while the kernel layer manages core functionalities and resources, including scheduling, context management, memory management, storage management, and access control. The hardware layer handles physical computing resources. Key contributions of AIOS include: - **New Agent-serving Architecture**: AIOS separates agent applications and resources into distinct layers, facilitating systematic resource management and optimization. - **AIOS Kernel Design and Implementation**: The kernel includes specialized modules for handling agent requests, scheduling, memory management, storage management, tool management, and access control. - **AIOS SDK Development**: The SDK provides a higher-level abstraction of kernel functionalities, allowing developers to focus on application logic without worrying about complex system calls. - **Empirical Results**: Experiments demonstrate that AIOS maintains or enhances agent performance on standard benchmarks and significantly improves execution efficiency, achieving up to 2.1× faster execution for various agent frameworks. The paper also discusses related work and concludes by outlining future directions for refining and expanding the AIOS architecture to better address the evolving needs of LLM-based agent development and deployment.The paper introduces AIOS, an innovative operating system designed to manage and optimize the deployment of LLM-based agents. AIOS addresses the challenges of resource management and scheduling in agent systems, which can lead to inefficient or harmful resource allocation. The architecture of AIOS is divided into three layers: the application layer, the kernel layer, and the hardware layer. The application layer provides an SDK for developers to build and integrate agents, while the kernel layer manages core functionalities and resources, including scheduling, context management, memory management, storage management, and access control. The hardware layer handles physical computing resources. Key contributions of AIOS include: - **New Agent-serving Architecture**: AIOS separates agent applications and resources into distinct layers, facilitating systematic resource management and optimization. - **AIOS Kernel Design and Implementation**: The kernel includes specialized modules for handling agent requests, scheduling, memory management, storage management, tool management, and access control. - **AIOS SDK Development**: The SDK provides a higher-level abstraction of kernel functionalities, allowing developers to focus on application logic without worrying about complex system calls. - **Empirical Results**: Experiments demonstrate that AIOS maintains or enhances agent performance on standard benchmarks and significantly improves execution efficiency, achieving up to 2.1× faster execution for various agent frameworks. The paper also discusses related work and concludes by outlining future directions for refining and expanding the AIOS architecture to better address the evolving needs of LLM-based agent development and deployment.