A new position paper argues that AI agents need something computers got decades ago: an operating system.
Right now, every agent framework builds its own version of the same plumbing: memory, spending limits, safety checks, and verification, even though protocols like MCP and A2A already make it easy to connect tools and agents. A paper posted to arXiv on September 18, 2026 proposes fixing that with a "Foundation Model Operating System," or FMOS, a layer that virtualizes access to foundation models the way a virtual machine abstracts physical hardware. Internally, the FMOS would manage memory tiers, choose which model handles a task and how many resources it gets, and enforce policy and verification rules. The authors compare its behavior to a brain switching between fast intuition and slow, deliberate reasoning, stepping in only when needed and otherwise adapting its own policies from experience.
The pitch lands because the problem is real and familiar. It echoes the fragmentation that plagued early computing before operating systems standardized basic services, and it is why building multi-agent systems today still feels like assembling a car from parts that do not share a chassis.
This is a proposal, not shipping code, so the real test is whether any lab or open-source project actually builds the FMOS layer, and whether developers adopt a shared substrate instead of their own bespoke runtimes.