Researchers propose AgentOS, a new way to design personal computers where people talk or type naturally and a central “Agent Kernel” carries out tasks. Instead of the familiar desktop with icons and windows, AgentOS would offer a single natural language or voice portal. The authors argue this change is needed because modern autonomous agents built from Large Language Models (LLMs) do not fit well into today’s Graphical User Interface (GUI) operating systems.

The paper builds on recent examples such as OpenClaw, an open-source, locally hosted agent that can read and write files, run terminal commands, manage calendars, and browse the web. The authors sketch an architecture with three main parts. First is the Single Port: a persistent multimodal interface that accepts voice, text, and context signals. Second is the Agent Kernel: the system core that turns user intent into actions, keeps long-term context, and coordinates many smaller agents. Third are Skills-as-Modules: reusable, composable pieces of behavior that users can define in plain language and combine to automate workflows (for example, a rule that extracts invoice totals from email PDFs and drafts a payment authorization).

At a high level, the Agent Kernel has two complementary jobs. Its northbound side does continuous intent parsing: it reads ambiguous human inputs and turns them into structured, machine-executable intents while maintaining conversational state and personal context. Its southbound side runs a Multi-Agent System (MAS) to split those intents into executable steps and send them to the underlying system using a Model Context Protocol (MCP) that talks to the legacy operating system components (files, network, drivers). The kernel also must schedule scarce LLM resources—things like context window size, token budgets, and API rate limits—so many concurrent agent threads can run without hitting memory or throughput limits.