Designing the AI-Native Operating System

Research question

What changes when an operating system schedules probabilistic inference requests across heterogeneous models and accelerators rather than deterministic threads?

Proposed architecture

The paper presents SlyOS as a layered reference architecture: agentic applications, a cognitive scheduling plane, an inference runtime, heterogeneous edge accelerators and optional cloud capacity.

Scheduling model

Inference placement is framed as continuous multi-objective optimization under latency, energy and integrity constraints, including intermittent connectivity and deterministic fallback.

Memory and trust

The work identifies key-value cache management as a central constraint and argues for continuous runtime attestation across heterogeneous inference paths.

Limits and next evidence

This is a preprint and reference architecture, not a claim of a completed client deployment. Comparative benchmarks, implementation evidence and independent review remain appropriate next steps.