Position Paper: Why Foundation Models Need a Virtualizing "Operating System" Layer

It is a proposed system layer sitting between applications and foundation models/agents, providing unified memory, automatic model selection, resource allocation and policy enforcement, so applications see a trusted virtual model instance rather than calling models directly.

MCP and A2A solve connectivity between tools and agents. They do not solve state, memory, budgets and guardrails, which each framework still rebuilds internally. FMOS targets that missing runtime layer above connectivity.

No. It is a position paper offering an architectural argument and a research call, with no benchmarks or empirical comparisons, so it is not evidence of deployment readiness.

Not yet. The practical move is to follow the debate and design current systems with portable, standardized memory and policy layers until reference implementations and standards mature.