Why AI Agents Need an Operating System of Their Own

Brian Chesky has spent the past year insisting a chatbot is the wrong interface for commerce. This week the Airbnb co-founder and CEO went further: the industry's entire approach to AI agents is built on a flawed foundation. Agents are being bolted onto operating systems that were never designed for them, and until that changes, consumer AI will not work.

"In an ideal world, AI would be operable at the kernel level," Chesky told TechCrunch. "Right now people are building AI apps for iOS, macOS, and Windows that are not AI operating systems."

The diagnosis lands at a moment when the gap between the promise of autonomous agents and their real-world performance has never been wider. Consumer agents can book a flight in a demo but fall apart the moment a task requires identity, payment, comparison, or collaboration. The problem, Chesky argues, is not the models. It is the substrate underneath them.

The Kernel Problem

Today's agents run as ordinary applications on top of an operating system that treats them like any other program. That forces every agent to re-implement the same primitives from scratch: memory, identity, permissions, scheduling, and a way to reach other software. There is no shared runtime, no standard SDK, and no interoperability layer beneath the surface.

Chesky points to OpenAI's early attempt at a ChatGPT app store as the cautionary tale. "If you want to be an app store like the iPhone, you need to have a software developer kit and an operating system like the App Store," he recalled telling Sam Altman. "Otherwise, none of our functionality is going to exist. And of course, it didn't work very well."

The same pattern is repeating. Every frontier lab is racing to become the quarterback, the single primary agent users route everything through, while nobody is building the underlying operating system that would let agents actually interoperate. It is a race to own the top of the stack while the bottom of the stack stays empty.

From Apps to Agents

Chesky's framing is useful for builders. Agents, he says, are "apps with faces." The difference is that apps have never been interoperable unless two companies struck a business deal, whereas agents, when they can talk to each other, make every app interoperable by default. He envisions the Airbnb app itself becoming a macro-agent, with the explore tab, customer service, and other functions each running their own agent, all connected through MCP, the Model Context Protocol championed by Anthropic.

The implication is that an agent's value lives not in what it can do alone but in what it can reach. A single agent locked inside one vendor's ecosystem is just an app with extra steps. The winning platform will be the one that gives agents a common substrate for talking to each other, not the one that captures the most single-purpose agents.

Chesky tested this himself: "I've used Airbnb on Muse and Instinct, and it doesn't work very well." The reason is not model quality. It is that those agents have no native way to reach Airbnb's identity, maps, payments, and messaging, and no rich interface to hand the user when a chatbot alone is not enough.

The Race for the Agent OS

The pieces of an agent operating system are already being assembled, just not by the companies you might expect. MCP supplies the interoperability layer. OpenClaw, Peter Steinberger's open-source operating system for agents, supplies a runtime. Nvidia's Open Agent Safety Platform pairs the OpenShell software boundary with hardware-level monitoring through its BlueField DPUs, giving agents an isolated execution environment. The major clouds, meanwhile, are racing to build sandboxes for agent workloads.

What is missing is the kernel: the layer that ties memory, identity, permissions, and scheduling into a single coherent runtime. Every agent today manages its own memory. Every agent negotiates its own identity and permissions. Every agent schedules its own work. That duplication is the single biggest source of fragility in production agent deployments, and it is exactly the layer Chesky is pointing at when he says AI should operate at the kernel level.

What This Means for Builders

For teams shipping agents today, the critique carries concrete consequences. First, treat interoperability as a first-class requirement rather than an afterthought: an agent that cannot reach other tools and other agents through MCP is already obsolete. Second, stop re-implementing memory and identity per agent, and use shared primitives wherever they exist. Third, do not bet everything on a chat-only future. Chesky's insistence that apps will not become "data layers of a universal simple UI" is a warning that the interface still matters, and that a designer can prompt better software than a consumer can.

The deeper lesson is that the bottleneck has shifted. A year ago the scarce resource was model capability. Today the models are good enough, and what is scarce is the runtime around them. Whoever builds the operating system that agents actually run on, whether Apple, Google, an open-source project, or a company that does not exist yet, owns the next platform shift the way iOS and Android owned the last one. Chesky's bet is that nobody has cracked it, and that is precisely the opportunity.