Brian Chesky: No Mainstream AI Without a Real OS for Agents

Chatbots still aren't the right interface

Airbnb has long been cautious about AI. Its new AI-powered search, shipped with the fall update, doesn't change a long-held conviction of Brian Chesky's: chatbots aren't built for discovery or purchasing.

His argument is practical. A chatbot offers a few options at a time and requires several exchanges before reaching a conclusion. For a flight, users want speed. For a place to stay, they like to browse.

Chesky points to studies showing people get more enjoyment from planning and anticipating a trip than from living it. Removing inspiration and browsing would remove part of the product. He envisions an in-between interface, somewhere between the chatbot and the current version.

Multiplayer AI, and screens that are half fixed, half generated

The second flaw of chatbots according to Chesky: they're built for one person. But Airbnb is often used by families or groups of friends. Hence the priority for the next three to six months: "multiplayer" AI, where several people interact at the same time.

On interface design, he doesn't buy into the fully generative vision pushed by startups like Monogram or Wabi, which build screens on the fly. He imagines a mix of deterministic interfaces, designed in advance, and AI-generated interfaces.

Making Airbnb legible to agents

Facing consumer agents like Instinct or Muse, which act on the user's behalf, Airbnb wants to strengthen its software infrastructure and become "agent-friendly." Chesky doesn't see these agents as a threat to the customer relationship. He views them more as good lead generators, something chatbots have already proven, in his view.

He refuses, however, to let apps become mere data layers behind a universal interface. Messaging with the host, comparison shopping, identity verification, maps: all of this requires either handing off to the app or a far more robust SDK on the agent side.

Down the line, he expects every part of the app to have its own agent: exploration, customer service, other features. An Airbnb "macro-agent" will eventually oversee them and talk to other agents via MCP. To him, agents are "apps with a face" that make services interoperable without a prior business deal.

The real problem: no AI-native OS exists

The core of his thesis lies elsewhere. Today AI apps are built for iOS, macOS, or Windows, none of which are operating systems designed for AI. Ideally, AI would operate at the kernel level, and all agents could interact with one another.

Chesky sees players fighting to become the "quarterback," the main agent everything flows through, without anyone building a complete SDK. He says he warned Sam Altman that the app store launched by ChatGPT wouldn't work without an SDK or a real OS. According to him, it indeed hasn't worked well.

His verdict is blunt: he tested Airbnb through Muse and Instinct, and the result is bad, worse than for hotels. In his view, until Apple, Google, or someone else builds a new platform, we'll have apps that contain agents, not apps that have become agents.

Two misconceptions, voice, and fewer meetings

Chesky flags two widespread misconceptions in Silicon Valley. The first: believing the interface will disappear in favor of speech and text. The second: believing software will become entirely generative. In his view, a designer will always get better software from a prompt than a consumer will.

He does believe, though, that we'll eventually talk to our machines more than we type. Airbnb will roll out voice agents this fall for search and customer service.

Internally, AI supposedly lets the company ship features faster. On a personal level, Chesky explains that it reduces "communication layers": he gets information by querying his AI rather than multiplying meetings.