Position Paper Proposes a Self-Evolving Operating System Layer for Foundation Models

Summary

AI applications are moving from single, monolithic foundation models toward compound agentic systems, but the software stacks supporting them remain fragmented. Although protocols such as MCP and A2A improve tool and agent connectivity, individual frameworks still embed their own runtimes for state, memory, budgets, and guardrails. The position paper compares this situation with computing before operating systems, when each application reimplemented basic services. It proposes a Foundation Model Operating System, or FMOS, as a system layer that virtualizes interactions with foundation models in a way analogous to virtual machines abstracting physical hardware. Applications would interact with what appear to be dedicated, trustworthy foundation model instances with effectively unbounded capabilities, while FMOS manages the underlying complexity. Its proposed functions include coordinating knowledge across multiple memory tiers, selecting models, allocating resources, and enforcing verification and policies. The paper also envisions FMOS learning when to intervene and when to let inference proceed directly, using operational experience to continuously adapt its policies. The article presents this as a position for a more portable and governable agentic AI stack, rather than reporting an implemented system or experimental validation.