Born from research

Research

In 2024 we were the first to propose the agent-centric operating system. While the industry was adding assistants on top, we set out to redefine the operating system itself. Kleinia is that vision, built in full, and our research keeps moving ahead.

Approaches

Three ways to bring agents to the desktop

Each is a sound choice for a different goal. Kleinia takes the third, because it is the only one that lets agents reach every part of the system.

01
Assistant app
A chat or copilot app on an existing OS. Quick to adopt and easy to replace.
Sees what the OS and other apps choose to expose.
02
Agent runtime on the OS
A framework that schedules and coordinates agents in user space, on top of an unchanged kernel and services.
Manages agents well, but the OS beneath stays unaware of them.
03
Agent-native OS
The operating system itself is built around agents: its services are agents, and it provides agents with identity, context and permissions.
Agents reach the whole system, and the system can improve itself.
Kleinia

Design principles

What we build Kleinia to do

The principles Kleinia is designed around, and how each one shows up in the system.

Agents are first-class citizens
The OS manages agents the way it manages processes, giving each one an identity, quotas, permissions and context. Your favourite agents and apps need no changes: they are taken in the moment they start.
Anything as an agent
Apps, services, login sessions and system components all exist as agents, with one identity, one capability description and an agent:// address. Even apps never adapted for AI can be found and called.
One model gateway
A single endpoint carries every model call. Keys, model permissions and quotas are managed by the system, not by each agent on its own.
The system, scheduled by agents
A resource agent understands what you are doing and picks the best CPU, I/O and page-cache policies, switching as soon as the load changes.
Intelligence apart from the fast path
Agents recognise and decide; the kernel runs only deterministic, safety-checked policies. Policies load and switch at runtime, and fall back to the default if anything goes wrong.
Memory that grows like a mind
The system keeps context, so agents carry project conventions and past decisions across sessions. Knowledge is not just stored but updated, evolved and reorganised.
A system that learns you
Faced with unfamiliar loads, the scheduling model evolves in the background, and devices share what they learn, so a new machine does not start from zero.
In control, in full view
Tiered permissions: routine actions run on their own, sensitive ones notify you, critical ones wait for you. Every model call, tool call and system action in a task can be traced.

Our research

The papers our team wrote, from the architecture Kleinia is built on to the agents that run inside it.

arXiv:2411.177102024-11 · cs.OS

ACOS

Agent Centric Operating System — a Comprehensive Review and Outlook for Operating System

Shian Jia, Xinbo Wang, Mingli Song, Gang Chen

Reviews the evolution of operating systems and proposes the Agent-Centric Operating System: every system component abstracted as an agent, connected through a flat Agent Communication Interface, under the principle “Agent for OS, OS for Agent”.

Foundational

The architecture of Kleinia itself — agent abstraction, ACI, system-service agents and agent-facing OS services.

→Agent abstraction: composable, instrumental, collaborative, scalable
→Flat architecture over the Agent Communication Interface (ACI)
→Four-property security model for user → agent → link → device

Since 0.1 AloeDefaultPaper ↗

arXiv:2511.116282025-11 · cs.DC

LAPS

Beyond the Monolithic Scheduler: Lightweight Adaptive Policy Switcher for Process Scheduling Optimization

Xinbo Wang, Shian Jia, Ziyang Huang, Jing Cao, Mingli Song

A Lightweight Adaptive Policy Switcher that recognises the current workload and switches the Linux process scheduler to the best-suited policy at runtime, instead of relying on one monolithic scheduler.

In Kleinia

Powers the Compute Scheduler Agent: the desktop picks the right scheduler for what you are doing, live.

→8.91% average performance gain over EEVDF
→Up to 91.7% lower latency in latency-sensitive tasks
→Open source

IntegratingSource ↗Paper ↗

arXiv:2510.159662025-10 · cs.AI

PISA

PISA: A Pragmatic Psych-Inspired Unified Memory System for Enhanced AI Agency

Shian Jia, Ziyang Huang, Xinbo Wang, Haofei Zhang, Mingli Song

A unified agent memory system inspired by Piaget: a trimodal schema mechanism (update, evolution, creation) and hybrid symbolic–neural retrieval, evaluated on LOCOMO and the new AggQA benchmark.

In Kleinia

Powers the Knowledge & Memory Agent: long-term, organised memory shared safely between agents.

→Schema update · evolution · creation
→Symbolic reasoning + neural retrieval
→Introduces the AggQA benchmark

IntegratingPaper ↗

Research we build on

Kleinia tracks the frontier beyond our own lab. When outside research makes the system better, we bring it in and credit it here.

SOSP 2025Peer-reviewed

cache_ext

cache_ext: Customizing the Page Cache with eBPF

Tal Zussman, Ioannis Zarkadas, Jeremy Carin, Andrew Cheng, Hubertus Franke, Jonas Pfefferle, Asaf Cidon

Linux uses one page-cache eviction policy for every workload. cache_ext lets custom eviction policies be written in eBPF and loaded per cgroup, without changing the kernel, in the same spirit as sched_ext for CPU scheduling. Policies such as MRU, LFU, S3-FIFO and LHD deliver up to 70% higher throughput and 58% lower tail latency.

Built on in Kleinia

Built into the Kleinia kernel alongside sched_ext and iosched_ext, so agents can take over page-cache eviction per workload.

→Up to 70% higher throughput
→Up to 58% lower tail latency
→Per-cgroup policies, no kernel changes

Since 0.1 AloeDefaultSource ↗arXiv ↗Paper ↗