Skip to content

Agent-centric operating system

Kleinia follows the agent-centric operating system we proposed in our survey ACOS (arXiv:2411.17710). This page summarises the parts that matter for using the system.

To a conventional operating system an agent is just a process: identity, context, permissions and resources are left to each application, and the system does not know agents exist. Meanwhile the system’s own scheduling policies are fixed at build time and cannot adapt to changing workloads.

ACOS answers both by making the agent the basic unit of the operating system: the system serves agents, and agents help run the system.

Direction Meaning In Kleinia
Agent for OS Agents schedule and maintain the system Adaptive scheduling, an extensible scheduling substrate, self-evolution, shared experience
OS for Agent The system serves agents Zero-modification onboarding, model gateway, one address space, context and memory, observability

Each is the condition for the other: because the system gives agents identity, permissions and an audit trail, they can safely take part in scheduling it; as they improve the system, every agent running on it benefits.

ACOS describes four properties of agents:

  • Composable: agents combine into larger agents or split into smaller ones.
  • Instrumental: one agent can serve as another’s tool.
  • Collaborative: agents exchange messages, divide work and resolve conflicts.
  • Extensible: new agents join through the same interface as existing ones.

Kleinia is named after the Klein bottle, a surface with no inside and outside. The system and its agents are the same: each holds the other.

© 2026 The Kleinia AuthorsStarted and maintained by the VIPA Lab