# An AI Operating System, or a Toll Booth *The next AI platform is a coordination layer. Everything depends on whose machine it runs on.* *By Shirish Garg · August 30, 2026 · Cognatu* *Canonical URL: * *"Plan a three-day trip to Tokyo for my family."* Nine words go into a box. Six systems wake up. A small model works out what you actually meant. A search index goes looking for what is true this week rather than what was true in training. A reasoning model builds an itinerary. A booking agent checks what is available. A payment rail waits its turn. And somewhere in there, something supplies the part you never typed: that you hate connecting flights, and that "my family" contains a seven-year-old who will not walk four miles a day. You picked none of them. You picked nothing at all. ![One typed request, "Plan a three-day trip to Tokyo for my family", fans out to six systems: a small model parsing intent, a search index, a reasoning model, a booking agent, a payment rail, and a local store of personal context. All six return into a single answer. The user picked none of them.](https://cognatu.com/blog/assets/coordination-layer.svg) *The user made one decision. The layer made six.* That is the whole trick, and almost nobody is arguing about who gets to perform it. We are arguing about models instead. Which one is best. GPT or Claude, Gemini or Llama, bigger or smaller. It is the wrong question and it is getting wronger every quarter, because models are becoming cheap, capable and easy to swap. Developers already route different tasks to different ones, and the plumbing that lets AI systems reach tools is being standardised in the open[1]. The fight has moved up a floor. To whoever decides which model gets used, what context it receives, which tools it can reach, and where the result goes. That layer is going to look a lot like an operating system. The question worth arguing about is whose machine it runs on. ## The operating system was never the interface Most people picture Windows or macOS as the thing on the screen. It isn't. The desktop is the interface. The operating system is the part underneath, deciding what gets access to what, letting thousands of applications share one machine without any of them having to negotiate with each other. Nobody thinks about it. That was the point. It also created some of the largest businesses in history. AI has produced the same problem wearing new clothes. One task wants a small fast model, another wants deep reasoning, another wants your calendar. There is not going to be one model that is best at everything, which means something has to sit above all of them and choose. The intelligence is fragmented. The experience isn't. That gap is the job. ## Your camera is better protected than your life Routing is only half of it. The layer also has to know who you are. I made the case for holding your own personal context in [the vault post](/blog/your-life-shouldnt-have-to-log-in) and won't rerun it here. What the operating system framing adds is a better word for what context is: a *resource*. Not a feature an app ships. Not a database a company keeps. Something the system manages on your behalf and lends out one request at a time. Which is exactly how your camera already works. An app doesn't own it. It asks, you approve, the grant is scoped, and you can take it back on a Tuesday afternoon for no reason. Nobody finds this exotic or difficult or a threat to innovation. It is simply the resource layer doing its job. Your camera has that. The accumulated record of how you think does not. ![Two rows drawn with identical geometry. In the top row an app reaches the camera only through a permission that is asked for, scoped and revocable, with a dashed arrow showing it can be taken back. In the bottom row the same path from your life to three apps runs through an empty dashed slot labelled no layer here, and each app ends up holding its own permanent copy.](https://cognatu.com/blog/assets/context-vs-camera.svg) *The difference is a box that doesn't exist yet.* Capabilities land in the same place. I argued in [the last post](/blog/ai-shouldnt-be-somewhere-you-go) that skills rather than apps become the unit of intelligence. In operating system terms that is an old shape: a capability is a service the layer can call, and the application goes back to being where the work happens rather than the thing that owns the ability to do it. ## Why the layer is worth so much money Platforms win by controlling distribution. Windows distributed software, the App Store distributed apps, search distributed information. The next one distributes intelligence. And distribution matters precisely *because* the models are commoditising, which was [the argument two posts ago](/blog/the-token-is-the-new-bandwidth). If ten models do a job at roughly equal quality, you genuinely do not care which one answered. The layer cares enormously. It is the only party in the transaction that knows what everything costs. Which produces the sentence this whole post exists to set up: **The model gets paid for inference. The layer captures the relationship.** That is the prize. It also explains the behaviour of every large AI company considerably better than anything they have said about benchmarks. ## Which is exactly why it matters where it runs Here's the part most versions of this argument skip. If the coordination layer lives on somebody's servers, it isn't an operating system in the honest sense. It's a toll booth. Everything you ask, every piece of context you accumulate, every capability you reach for passes through a company with both the incentive and the position to take a cut and keep a copy. "Whoever owns the layer wins" is only a thrilling sentence if you are the one who owns it. For everybody else it is a sentence about a tax. An operating system, in the original meaning, works for the person holding the machine. It manages resources on your behalf. It enforces permissions in your interest. It is the reason an app can use your camera without owning your camera. Applied here, that means the layer runs on your device. Your context stays local and goes out one scoped, revocable request at a time. Models are things the layer calls rather than empires you live inside. Routing decisions get made in your interest, not in the interest of whoever has the best margin on inference this quarter. None of this is waiting on a research breakthrough. Apple exposes the on-device model behind Apple Intelligence to developers directly[2]. Android reaches Gemini Nano through ML Kit without the text leaving the machine[3]. The tool-calling layer above them is being standardised in the open[1]. Both versions of this thing are buildable now, out of parts that already shipped. ![Two architectures drawn side by side. On the left, every request and every piece of context routes through a remote coordination service before reaching any model. On the right, the coordination layer sits on the device; context never leaves it, and models are called out to as interchangeable services.](https://cognatu.com/blog/assets/toll-booth-vs-os.svg) *Same capability. Opposite defaults about who holds the context.* One of them reruns the last twenty years with a more intimate data stream. The other is the first realistic chance to build the intelligence platform without the surveillance layer underneath it. I should say plainly that we are not neutral about which one arrives. Cognatu is being built for the second, so treat this as an argument I have a stake in rather than a forecast I am reporting. I could also be wrong about the timing. On-device coordination has been six months away for a while now, and I have no special insight into when that stops being true. ## The real platform war We have seen this film. Microsoft didn't write every application. Apple didn't build every iPhone app. Google didn't create the web. Each one won by becoming the layer everything else had to stand on. AI is growing another one right now, in public, while the commentary is still busy scoring benchmarks. The only question left is whether it ends up as a layer everything passes *through*, or a layer that works *for* the person on the other side of it. Those are not the same platform. From the outside they will look identical for a while, which is the difficulty. **Unlike last time, we can still tell them apart before it's built.** --- **See what leaves the device** Our demo runs a real conversation, works out what it is about, and matches an ad without the text ever leaving the machine. Open your network tab and watch what crosses. That is the same constraint this post is asking of the coordination layer. No account needed. [Watch it run](https://cognatu.com/demo) · [How the SDK works](https://docs.cognatu.com) *Next: [AI Shouldn't Be Somewhere You Go](/blog/ai-shouldnt-be-somewhere-you-go). Where the layer actually touches you, and why it is the text field you are typing in right now.* --- **Sources** 1. Model Context Protocol, [Introduction](https://modelcontextprotocol.io). An open standard for connecting AI systems to tools and data sources, so the tool layer is not owned by a single vendor. 2. Apple Developer, [Foundation Models framework](https://developer.apple.com/documentation/foundationmodels). Direct access to the on-device model behind Apple Intelligence, offline, with nothing sent off the device. 3. Google, [ML Kit GenAI APIs](https://developers.google.com/ml-kit/genai). Summarisation, proofreading and rewriting on-device; input, inference and output never leave the machine. *Cognatu builds privacy-first infrastructure for AI apps · [cognatu.com](https://cognatu.com)*