AI applications are moving from single, monolithic foundation models to composite agentic systems. Today’s stacks remain fragmented: protocols such as MCP and A2A ease tool‑agent connectivity, yet each framework embeds an implicit runtime for state, memory, budgets, and guardrails, making behavior non‑portable and governance brittle. This mirrors the pre‑operating‑system era when every program re‑implemented basic services. This position paper proposes a Foundation Model Operating System (FMOS) that abstracts FM interactions like a virtual machine abstracts hardware, giving applications the illusion of dedicated, trustworthy FM instances with effectively unbounded capabilities. Internally, FMOS orchestrates multi‑tier memory management, model selection, resource allocation, and verification with policy enforcement. It learns when to intervene and when to let inference run directly, continuously adapting its policies based on operational experience, much like the brain switches between fast intuition and slow deliberation.
Review