Researchers have proposed an operating system called CrabOS as a shared work environment in which people and AI can take turns continuing the same task. The authors frame the idea as moving support for that handoff from application-level bridges to the operating-system layer.
The work is a system proposal, not a user trial. The reported evaluation is a capability-based architectural comparison with mainstream agent systems, supported by descriptions of the design and qualitative case studies. No participant or dataset sample is described.
The handoff is built into the design
CrabOS is built around three design principles. It uses stable text objects that can be read in natural language; unified, referable work objects that give humans and AI the same items to work with; and the same auditable capability entry for every executor. In plain terms, the proposal aims to make task state readable, addressable and subject to a common record of how actions are invoked.
Those principles are arranged across four layers: the L0 Object Layer, the L1 System Services Layer, the L2 Shell Layer and the L3 App Layer. The architecture separates objects, services, shells and applications, giving the shared workspace a defined structure.
At the object layer, CrabOS makes a distinct architectural choice. It does not restrict access to system state through a query protocol. Instead, it materializes every system object as a text object and uses memory only as a cache. The design therefore treats system state as a readable object that can be referred to directly, rather than only as information retrieved on request.
Cases built from the same objects
The memory-management case study applies this model to an L3 app. It says any memory strategy can be implemented independently as an application, treating memory management as a separately defined part of the operating system's app layer.
The context case is more concrete. A PDF viewer's runtime state is represented as an L0 object. When the reader changes pages, that change is written immediately to an addressable object that an AI agent can read. The proposed result is a handoff in which the agent can locate the human's current place in the document through shared state.
CrabOS uses the same logic for orchestration. The case models a Task as an L0 text object and says different workflow styles can be composed from existing object primitives, rather than requiring a new mechanism for each orchestration pattern.
The operation-trace case routes capability invocations from every L3 app through Kernel Interface tool gating, a common gate for tool calls. When those invocations are persisted, the resulting traces are recorded at one semantic-operation granularity, meaning the trace is organized around a single meaningful system action. The supplied analysis does not report an empirical assessment of how complete or useful those traces are.
What remains untested
The design also draws a boundary around the existing software environment. CrabOS has no built-in arbitrary Bash execution because shell execution could bypass the Kernel Interface; instead, it uses SSH to interact with an external execution environment. Existing native applications must first be migrated into an L3 app to participate in CrabOS work objects, although a GUI adapter can leave the native backend unchanged.
That architecture brings an I/O trade-off. Persisting the complete L0 Object Layer introduces more write latency than an in-memory database. CrabOS describes delayed writes for high-frequency streaming output, but the paper does not report a quantitative result for that trade-off.
The paper therefore leaves its main promise open. Its architecture provides a proposed way to share state, work objects and capability records between human and AI executors, but the evidence does not show lower handoff effort, fewer errors, greater reliability or better performance than existing systems. Those questions would require comparative workload studies and human or system evaluations.
The manuscript is an arXiv version 1 preprint dated 28 Aug 2026. Its supplied front matter lists a Peking University affiliation and no funding statement.
Paper data and sources
Original title: CrabOS: An Operating System for Human-AI Co-inhabitation
Authors: Qi Yang, Yun Ma
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text