FMOS academic proposal invokes computing heritage to tame agent chaos

Academic researchers presented a scientific position paper at the position papers track of ICML 2026, calling for a radical restructuring of AI software architecture by establishing an independent operating layer called the Foundation Model Operating System (FMOS). The paper, prepared by ten researchers including Subarna Bhattacharya, Tarun Kumar, and Ian Foster, starts from the fact that smart applications have practically shifted from relying on a single stacked foundation model to composite agent systems that combine multiple models and software tools, a change accompanied by significant fragmentation in currently used software packages.

The paper notes that emerging protocols such as MCP and A2A, despite their success in facilitating communication and integration between agents and external tools, have not resolved the deep fragmentation problem. Each software framework embeds its own implicit runtime environment to manage transient state, memory, consumption budgets, and strict security controls. This internal variation makes system behavior non-portable across platforms and renders governance and monitoring processes vulnerable to brittleness and failure. The researchers compare this situation to early-era computing before operating systems, when every application program was forced to rebuild and manage core services itself.

The paper proposes addressing this dilemma by creating a FMOS layer that virtualizes model interaction, echoing the historical role of virtual machines in abstracting and isolating computer hardware. This architecture gives software applications the impression that they are dealing with customized, reliable instances of models with virtually unlimited capabilities. The proposed operating layer performs several concurrent functions from within:Coordinating knowledge across multiple memory storage layersautomatic routing to select the most suitable model and efficiently allocate compute resources, alongside continuous verification of operations and centrally enforced usage and security policies.

The proposed system architecture draws inspiration from the human brain, specifically the flexible switching between rapid automatic intuition and deliberate deep thinking. Instead of constant costly administrative intervention, the operating system learns autonomously when to intervene to apply controls and audit pathways, and when to allow inference processes to proceed directly without delay, continuously evolving its internal policies based on the cumulative operational experience generated by actual use.

This structural shift directly affects tech companies and development teams in the Gulf and Egypt, especially financial institutions and government bodies that build bespoke agent systems for transaction services and sensitive data. Currently, local engineering teams must craft costly ad-hoc solutions for each AI system to manage inference and memory consumption and ensure compliance with sovereignty and local data regulations. Consolidating these services within an independent operating system layer reduces the financial waste associated with maintaining tangled code, eliminates reliance on specific frameworks, and redirects regional software talent toward policy and governance engineering instead of building unstable runtime environments.