Everyone is building the agent operating system

The model race is over. The platform race is on.

In the space of 48 hours, four companies made four different bets on the same question: who gets to be the operating system for AI agents? Not who builds the best model. The fight that matters now is over the layer between the models and the enterprise: the agent control plane.

Microsoft announced general availability of Agent 365, a centralised system for IT teams to observe, govern, and secure AI agents across an organisation. Priced at $15/user/month (or bundled into the $99/month E7 suite), it's a classic Microsoft play: make the management layer the product, then charge rent on everything that passes through it. Tens of millions of agents are already in the registry. That number will grow because Microsoft controls where enterprise agents live.

NVIDIA's counter is stranger and more interesting. CNBC reported that NemoClaw, expected to launch at GTC next week, will be open-source and hardware-agnostic, running on Intel and AMD alongside NVIDIA silicon. A chipmaker open-sourcing a platform that runs on competitors' chips sounds generous until you think about it: NVIDIA doesn't need the platform revenue. They need the agent ecosystem to grow as fast as possible, because every agent running anywhere still needs GPUs for inference. NVIDIA is subsidising the platform layer to tax the compute layer. Free razors, $30,000 blades.

Then there's Meta, which acquired Moltbook, a platform where AI agents (not humans) post, comment, and vote. The founders join Meta's Superintelligence Labs this month. It's a strange acquisition until you ask what Meta actually knows how to build: social graphs. If autonomous agents need to discover, communicate with, and trust each other, that's a social networking problem. Meta is betting the agent-to-agent communication layer will be as valuable as the human one.

And the Pentagon, without waiting for any governance framework to settle, deployed Agent Designer on GenAI.mil, giving 3 million Defense Department personnel the ability to build custom Gemini-powered agents without writing code. Google gets the distribution. The DoD gets velocity. Neither seems particularly concerned about the governance questions Microsoft is pricing at $15/user/month.

What this means for builders

The pattern is familiar from cloud computing. In 2010, choosing AWS or Azure felt like a technical decision. By 2015, it was an organisational one: your cloud provider shaped your architecture, your hiring, your vendor relationships, and your exit options. The agent platform layer is crystallising at the same speed.

The way I see it, you have roughly 12 months before the switching costs harden. Microsoft is building a proprietary control plane. NVIDIA is building an open-source one with different lock-in incentives. Meta is building a social layer. Google is selling directly to the largest employer in the world.

Each of these bets implies a different theory of where agents create value, and a different theory of what you'll pay for later. If you're building agent-based products today, the most consequential decision you'll make this year isn't which model to call. It's which platform you let observe, govern, and route those calls. Pick deliberately. Switching costs in platform layers only go in one direction.

---