Main Text

However, the increasing complexity of tasks, data volumes, and requirements for adaptability and personalization poses serious challenges to traditional OS based on predefined rules [3].

The convergence of MAS and AI has become especially relevant with the emergence of powerful Large Language Models (LLM), capable of understanding natural language, reasoning, and planning [2].

Various AI methods contribute to the creation of intelligent agents for operating systems [3].

Reinforcement learning (RL) becomes particularly important, especially in its multi-agent variant (Multi-Agent Reinforcement Learning, MARL) [4, 5]. MARL allows training a group of agents to work together to achieve common goals, such as optimizing resource allocation, planning task execution, or coordinating actions in a dynamic OS environment. Despite its promise, MARL faces challenges related to high-dimensional state spaces, environmental non-stationarity due to the actions of other agents, and scalability issues [4].

Large Language Models (LLM) have a revolutionary impact on the field. Due to their capabilities in understanding and generating natural language, planning, and reasoning, LLMs are becoming the foundation for a new generation of intelligent agents [1, 2]. In the context of OS, LLM-based agents can provide intuitive user interaction, assist in solving tasks, manage other tools, and form the basis for the concept of "Agent OS," where the agent system is deeply integrated into system management [1, 6].

A key shift in the field is the development of systems and frameworks that deeply integrate LLM agents into operating systems, transitioning from isolated intelligent functions to comprehensive system solutions [1, 6]. These architectures aim to solve resource management, efficiency, and coordination problems that arise when operating LLM agents.

Interaction between the AIOS Kernel and OS Kernel occurs through standard OS system calls, while interaction between applications and the AIOS Kernel occurs through specific AIOS system calls. The Hardware Layer includes physical components (CPU, GPU, memory, disks), access to which is mediated by the OS kernel. Experiments have shown that using AIOS can achieve significant acceleration (up to 2.1 times) when performing tasks with LLM agents under resource competition conditions [1].

Building on similar principles of resource management and task separation, the KAOS (Large Model Multi-Agent Operating System) architecture was also proposed – a multi-agent OS built on the open Kylin OS [6]. KAOS also uses a multi-level architecture with separation between the OS kernel and LLM kernel (see Figure 2).

The integration of AI and MAS into operating systems promises significant benefits, validated by the goals and capabilities of modern architectures:

• Increased efficiency and automation: Automation of routine tasks and optimization of workflows in OS [3].
• Optimization of resource management: Effective dynamic resource allocation, which is a key goal of architectures such as AIOS [1], especially for resource-intensive LLM agents.
• Enhanced user experience: Providing personalization, natural language interaction, and proactive assistance, as implemented in systems like KAOS using specialized agents and LLM [6].
• Improved reliability and security: The ability to proactively detect system failures and intelligent threat monitoring, which is an important aspect in developing stable agent systems [1, 6].

However, the implementation of such advanced systems comes with a series of serious challenges, actively addressed in current research:

• Coordination complexity and scalability: Managing the interaction of a large number of autonomous agents remains a difficult task [4], which systems like KAOS attempt to address using manager agents and vertical collaboration strategies [6].
• Resource requirements: The high computational cost of training and executing LLMs and other complex AI models [1, 6], an acute problem requiring efficient solutions at the OS kernel level, as proposed in AIOS and KAOS [1, 6].
• Privacy, security, and ethics: The need to protect user data used by agents, prevent malicious attacks, and ensure fairness of AI algorithms [3].
• Explainable AI (XAI): The difficulty of understanding and interpreting decisions made by complex models ("black box problem"), which is critical for user trust and system debugging.

The integration of artificial intelligence, especially large language models, into multi-agent systems fundamentally transforms approaches to the development and operation of operating systems. This enables the creation of more intelligent, autonomous, adaptive, and user-oriented computing environments. Advanced architectures such as AIOS and KAOS represent important steps toward realizing the concept of "Agent OS" [1, 6].

Список литературы (References)

1. Городецкий, В.И., Бухвалов, О.Л., Скобелев, П.О. Современное состояние и перспективы индустриальных применений многоагентных систем [Текст] / В.И. Городецкий... // Управление большими системами... – 2017.
2. Никитаев, В.М., Стаценко, В.Е. Применение многоагентной интеллектуальной системы мониторинга... // Вестник науки. – 2022.
3. Обзор и сравнение Multi-Agent Frameworks... // Хабр. – 2023. – Ссылка.
4. Large Language Model-Based Agents for Software Engineering: A Survey... // arXiv preprint. – 2023. – Ссылка.
5. [Здесь должен быть источник №5, упомянутый в тексте]
6. [Здесь должен быть источник №6 (KAOS/Agent OS), упомянутый в тексте]