AgentOS: Native LLM Agent OS

• AgentOS is a novel operating system paradigm that natively treats LLM-based intelligent agents as first-class entities alongside traditional processes.
• It integrates multi-modal observations and natural language–driven APIs to enable seamless human and device automation across desktop, web, and mobile platforms.
• The system employs robust resource scheduling, adaptive memory, and planning mechanisms to support autonomous, cross-domain task execution.

AgentOS is a new operating system paradigm designed for native hosting, orchestration, and lifecycle management of LLM–based intelligent agents. Contemporary AgentOS implementations repudiate the legacy division between process-centric resource management and GUIs, instead providing agents with direct, semantic, and multi-modal interaction capabilities over a device’s entire computational and interface substrate. Key design tenets include treating (multi-modal) LLM agents as first-class system entities alongside traditional processes; abstracting system state and user intent as LLM-consumable observations; embedding agent runtime, planning, grounding, and adaptive memory as core system services; and exposing natural language–driven APIs for user and agent control over workflow automation, resource manipulation, and environment orchestration (Hu et al., 6 Aug 2025, Mei et al., 2024, Liu et al., 9 Mar 2026, Li et al., 24 Feb 2026, Zhang et al., 20 Apr 2025).

1. Conceptual Foundations and Architectural Distinctions

AgentOS differs fundamentally from traditional operating systems in several domains:

• Agent-Native Process Model: Agents are managed directly by an Agent Scheduler and Agent Kernel with unified APIs for observation and action, placing LLMs and agent processes on equal architectural footing with threads and system services (Hu et al., 6 Aug 2025, Mei et al., 2024).
• Multi-Modal Observability: The observation space for agents natively includes pixel-level input (screenshots), structured environment representations (GUI trees, DOM, logs), and terminal/CLI state, facilitating robust grounding across desktop, web, and mobile environments (Hu et al., 6 Aug 2025, Zhang et al., 20 Apr 2025).
• Natural Language–Level APIs: AgentOS exposes system capabilities through intent-oriented, language-centric calls ("open email and reply to Alice") rather than only procedural syscalls (e.g., Win32 APIs), further mediated by built-in planners and grounding layers (Hu et al., 6 Aug 2025, Liu et al., 9 Mar 2026).
• Integrated Personalization and Memory: State, preferences, and past trajectories are stored hierarchically to enable agent self-evolution, long-horizon planning, and task contextualization at the system level (Hu et al., 6 Aug 2025, Li et al., 24 Feb 2026).
• Resource Control and Scheduling: AgentOS enforcement of strict resource allocation, preemptive context management for LLM inference, and per-agent access control prevents contention, runaway resource utilization, and Denial-of-Service conditions that are intractable in legacy OS/agent hybrids (Mei et al., 2024).

The principal motivation underlying AgentOS is the democratization and unification of device automation: enabling both technical and non-technical users to delegate arbitrarily complex, cross-domain automation tasks using natural language, overcoming the fragility and siloing of prior macros, RPA, or script-based solutions (Hu et al., 6 Aug 2025, Zhang et al., 20 Apr 2025, Liu et al., 9 Mar 2026).

2. Core Components and Data-Flow

The AgentOS stack is organized around a multi-layered architecture:

• Environment Layer: Abstracts the device’s interface surface—windows, files, keyboard/mouse for desktop; gestures and sandboxes for mobile; DOM/runtime for web (Hu et al., 6 Aug 2025).
• Observation Layer: Agents query an Observation object $Obs_t$ containing:
  • Screenshot image $I_t \in \mathbb{R}^{H \times W \times 3}$
  • Accessibility/DOM tree $T_t$
  • CLI prompt $C_t$
  • Status logs $L_t$

These are tokenized by vision/text encoders into a state vector $S_t$ (Hu et al., 6 Aug 2025).

• Action Layer: A superset action space with input (click, type, swipe), navigation (scroll, back), and extended operations (scripts, API calls, subagent spawning) (Hu et al., 6 Aug 2025).
• Reasoning and Planning: A core multi-modal LLM receives $S_t$, applies perception encoding and planning (Chain-of-Thought, ReAct), emits a task-level plan $\pi$, and delegates each plan step to a grounding/execution module determining system-level actions $A_i = g(w_i, S_t)$ (Hu et al., 6 Aug 2025, Zhang et al., 20 Apr 2025).
• Memory Subsystem: Persistent, hierarchical, and multimodal (action history, screenshots, user profiles, SOPs, vector-based retrieval), supporting continual learning and personalization (Hu et al., 6 Aug 2025, Li et al., 24 Feb 2026, Liu et al., 9 Mar 2026).
• Agent Scheduler and Kernel: Orchestrate agent lifecycle, context switching, preemption, and enforce per-agent privilege isolation and resource sharing (Mei et al., 2024, Zhang et al., 20 Apr 2025).

A high-level agent control loop can be formalized as:

loop: Obs_t ← AgentOS.get_observation() S_t ← PerceptionEncoder(Obs_t) π(·) ← LLM.plan(S_t, Goal) for each step in π: A ← GroundingModule(step, Obs_t) AgentOS.execute(A) Obs_{t+1} ← AgentOS.get_observation() if Goal achieved or timeout: break

[(Hu et al., 6 Aug 2025)