Answers are grounded in this paper and cite their sources.
An operator should be able to authorize a bounded task once, then let an AI agent complete it without repeated permission prompts. That requires more than a convenient tool interface: delegated authority must remain bounded, resource spending must survive worker failure, and a lost reply must not turn a retry into a duplicate effect. We argue for an operating-system contract organized around jobs whose authority, budgets and operation outcomes persist independently of their workers. OSIOS is an early standalone operating-system project exploring this contract through a capability-based microkernel and resource-owning services. The central research question is not whether a new kernel can run an agent, but what an operating system adds beyond a well-engineered runtime on Linux. We propose a staged comparison of tuned Linux, the same Linux with the OSIOS interface, and an OSIOS guest with that interface. The evaluation separates useful task completion, containment and recovery, and explicitly permits the conclusion that Linux captures all practical benefit. The boot handoff has passed 24 of 24 checks on the tested revision; agent-performance and job-isolation results remain future work.
No takes yet. Share an insight, caveat, or question.
Yaron Torgeman (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: