The landscape of artificial intelligence is shifting from single, monolithic foundation models (FM) toward compound agentic systems. However, a new position paper suggests that today's development stacks remain fragmented. While emerging protocols ease connectivity between tools and agents, each framework still embeds its own implicit runtime for state, memory, budgets, and guardrails, leading to non-portable behavior and brittle governance.
The Case for Virtualization
The researchers argue that the current state of AI mirrors the era of computing before operating systems, when every program had to re-implement basic services. To solve this, they propose a Foundation Model Operating System (FMOS)—a system layer designed to virtualize FM interactions. Much like virtual machines abstract physical hardware, the FMOS provides applications with the illusion of dedicated, trustworthy FM instances with nearly unbounded capabilities.
Orchestration and Self-Evolution
Internally, the FMOS is designed to manage knowledge across various memory tiers, handle model selection, and oversee resource allocation and policy enforcement. A key innovation of the proposed FMOS is its self-evolving nature. Drawing a parallel to the human brain's ability to switch between fast intuition and slow deliberation, the FMOS learns when to intervene and when to let inference proceed directly, continuously adapting its internal policies based on operational experience.
