AI operating system is the phrase Airbnb chief executive Brian Chesky reached for when TechCrunch asked him why consumer AI still feels unfinished. In an interview with Ivan Mehta published on 1 October 2026, a day after Airbnb switched on AI search for US guests, Chesky argued that the industry is building agents on top of iPhones, Macs and Windows PCs that were never designed for them. Until someone builds a true AI operating system, he said, apps will carry agents inside them rather than becoming agents that work together.
The interview covers more than the AI operating system headline. Chesky explained why he still thinks chatbots are the wrong shape for shopping, admitted that Airbnb works badly inside today’s consumer agents, recalled a warning he says he gave Sam Altman about ChatGPT’s app store, and set out Airbnb’s plans for “multiplayer” AI and voice agents this autumn.
This article explains what Chesky said, what an AI operating system would actually need, who is closest to building one, and what the argument means for businesses that want their services to work with agents. For the product launch that prompted the interview, see our report on Airbnb AI search and the fall update.
Table of contents
- What Brian Chesky Said About an AI Operating System
- Why Chesky Says Consumer AI Has Not Been Cracked
- The App Store Warning Chesky Gave Sam Altman
- What an AI Operating System Would Need
- Who Could Build an AI Operating System
- Generative Versus Designed Interfaces
- Multiplayer AI and Voice Agents at Airbnb
- How Airbnb Is Preparing for an AI Operating System
- What an AI Operating System Means for UK Businesses
- AI Operating System FAQ
- References and Further Reading
What Brian Chesky Said About an AI Operating System
Chesky’s argument starts with a complaint about where AI lives today. Companies are shipping AI apps for iOS, macOS and Windows, but none of those platforms is, in his words, an “AI operating system”. In an ideal world, he told TechCrunch, AI would be operable at the kernel level, the core of the operating system, and every component, including agents, would interact with every other.
The quarterback race
Instead, he sees a race to be the “quarterback”: the primary agent that a user goes through for everything. ChatGPT, Gemini, Meta’s Muse and a crowd of startups all want that seat. Chesky’s objection is that a quarterback without a full playbook cannot run the game. “The problem with that model is no one is developing a full operating system software developer kit,” he said.
Agents as “apps with faces”
His second idea is about interoperability. Apps, he pointed out, have rarely worked together unless two companies struck a business deal. Agents are different: he called them “apps with faces”, and argued that when agents talk to each other, every app can become interoperable. That is the promise an AI operating system would make standard rather than negotiated.
The macro Airbnb agent
Chesky also described where Airbnb itself is heading. “I think of the Airbnb app as one day becoming an agent,” he said, with an agent in the explore tab, another in customer service and a third being developed for another service area. “Probably eventually there’s going to be a macro Airbnb agent. And that will probably be interoperable with MCP to other agents.”
Why Chesky Says Consumer AI Has Not Been Cracked
The most striking line in the interview is an admission about Airbnb’s own experience inside other companies’ agents.
Airbnb inside Muse and Instinct
“I’ve used Airbnb on Muse and Instinct,” Chesky said. “And it doesn’t work very well. I mean, even hotels don’t work that well, but Airbnbs work even worse. So I don’t think we’ve cracked consumer AI.” Muse is Meta’s consumer agent, launched on 8 September, and Instinct is the viral agent startup that raised a $1 billion Series C at a $10 billion valuation, which we covered in our report on Instinct’s funding round.
His diagnosis is that the problem sits below the agents. “For consumer AI to be cracked, we have to crack at a lower substrate,” he said, adding that “all apps need to become agents; all must be interoperable; and all must have much richer user interfaces.” That lower substrate is what he means by an AI operating system.
Why holiday homes are harder than hotels
Chesky did not explain why Airbnbs fare worse than hotels, but the reason is easy to infer. A hotel room is close to a commodity: a date, a price and a room type. A home has a host, house rules, a unique layout, photos that do the selling and often a conversation before booking. An agent that flattens all of that into a text answer loses most of what makes the listing worth booking.
Two Silicon Valley misconceptions
Chesky named two beliefs he thinks are wrong. The first is that the interface is going away and people will simply talk and type. The second is that software will become completely generative. His counter-argument is that “a designer can prompt better software than a consumer can”, so carefully designed screens will keep their place.
Chatbots are still useful lead generators
Chesky has not changed his view that chatbots are poor for browsing and shopping, because they offer a few options at a time and need several turns to reach a result. But he conceded one point: fears that AI tools would cut companies such as Airbnb out of the customer relationship have eased, because chatbots have proved good at lead generation. Agents, he said, could play the same role.
The App Store Warning Chesky Gave Sam Altman
The interview’s sharpest anecdote concerns OpenAI. “ChatGPT, at the beginning of last year, tried to do their little app store,” Chesky said. “And I told Sam: 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. Otherwise, none of our functionality is going to exist. And of course, it didn’t work very well.”
What the timeline shows
Chesky’s date is loose. OpenAI’s app experiments did not happen at the beginning of 2025: the GPT Store opened in January 2024, and apps built with the Apps SDK arrived in ChatGPT in October 2025, with a directory following in December. Whichever attempt he meant, the substance matches his public record. In October 2025, when Booking.com and Expedia launched inside ChatGPT, Chesky told Bloomberg that Airbnb had stayed out because “I didn’t think it was quite ready”.
| Date | OpenAI app move | Airbnb’s position |
|---|---|---|
| March 2023 | ChatGPT plugins | Not a launch partner |
| January 2024 | GPT Store opens | No partnership announced |
| October 2025 | Apps in ChatGPT, Apps SDK | Stayed out: “not quite ready” |
| December 2025 | App submissions and directory | No change announced |
| 29 September 2026 | Extensions, Sign in with ChatGPT, Marketplace | Building its own AI search instead |
The gaps between attempts
OpenAI has returned to the idea repeatedly. Counting from each launch month to the next gives the intervals below: 10 months from plugins to the GPT Store, 21 months from the GPT Store to the Apps SDK, 2 months to the directory and 9 months to the DevDay 2026 additions.
Months between OpenAI’s app-platform launches (launch month to launch month)
What DevDay 2026 added
The latest version, announced on 29 September 2026, adds pieces an earlier list of apps lacked: a lasting interface inside the chat window, suggestions during a conversation, Sign in with ChatGPT and a marketplace for business software. We broke that down in our analysis of OpenAI’s app store model. It is closer to Chesky’s demand, but it is still a platform inside one app, not an AI operating system that every agent can share.
What an AI Operating System Would Need
Chesky did not publish a specification, but his comments add up to a list of what an AI operating system needs. The table compares each requirement with what exists today.
| Requirement | What Chesky said | Closest thing today |
|---|---|---|
| AI at the core | Operable “at the kernel level” | Assistants layered on iOS, Android, Windows |
| A full developer kit | “A full operating system software developer kit” | Apps SDK, App Intents, AppFunctions |
| Interoperable agents | Agent-to-agent, MCP | Model Context Protocol servers |
| Rich interfaces | “Much richer user interfaces” | Embedded app widgets in chat |
| Handoffs | Agent hands off or brings “the app to you” | Deep links back to the app |
| Group use | “Multiplayer” AI | Shared chats, still mostly one user |
Shared services, not separate silos
An operating system is valuable because it gives every app the same services: storage, notifications, permissions, payments and identity. An AI operating system would do the same for agents, so that an Airbnb agent could verify a guest’s identity, check a calendar and pay through a shared layer rather than through one-off integrations. Chesky’s list of features he cannot bring into a chat window today (“browse”, messaging the host, comparing, identity checks, maps) is a list of those services.
The handoff problem
Chesky described the choice that agent makers face. “Either the agent has to do a handoff or the agent application would need to build a bit more of, I think, a robust software developer kit to bring the app to you, so to speak.” A handoff sends the user back into the Airbnb app. A richer kit lets Airbnb’s interface appear inside the agent. An AI operating system would make the second option normal.
Permissions and trust
The hardest part of any AI operating system is deciding what an agent may do without asking. Operating systems already manage permissions for cameras, contacts and location. Agents need the same for spending money, messaging people and booking non-refundable stays, and no consumer platform has yet settled those rules across every app.
Who Could Build an AI Operating System
On who should build an AI operating system, Chesky’s answer was blunt: “It’s really up to Apple or Google, or somebody, to build a new platform for us to really make the true shift from apps to agents.” The contenders already hold pieces of an AI operating system.
Google and AppFunctions
Google is the furthest along in public towards an AI operating system for phones. Its Android developer documentation describes AppFunctions as a platform API that lets apps “behave like on device MCP servers”, exposing functions as tools for agents and assistants such as Gemini. It runs on Android 16 or later, callers need a dedicated permission, and Google says that as of May 2026 the Gemini integration is in a private preview with trusted testers. In February, 9to5Google reported the first supported devices as the Pixel 10 series and recent Samsung Galaxy phones.
Apple and App Intents
Apple’s route is App Intents, the framework it introduced in 2022 that lets Siri, Shortcuts and Apple Intelligence trigger actions inside third-party apps. It is a genuine operating-system layer, which is exactly what Chesky wants, but it was built for a voice assistant calling single actions rather than for agents negotiating with each other.
OpenAI, Meta and the assistant platforms
OpenAI, Meta and Google’s Gemini all control large assistant platforms, but none controls a phone operating system used by billions. That is why Chesky expects apps to remain central “as long as the iPhone remains dominant”. An assistant company can build a rich app platform inside its chatbot, yet it still runs as one app among many on someone else’s phone.
MCP as the shared wiring
Chesky expects the macro Airbnb agent to connect to others through MCP, the Model Context Protocol that Anthropic introduced in November 2024 and that has since become a common way for agents to reach software tools. MCP standardises the plumbing, not the interface, so it solves only part of an AI operating system on its own.
| Contender | Owns the device OS? | Agent hook today |
|---|---|---|
| Yes (Android) | AppFunctions, Gemini preview | |
| Apple | Yes (iOS, macOS) | App Intents for Siri and Shortcuts |
| OpenAI | No | Apps SDK, Extensions, Dots |
| Meta | No (glasses aside) | Muse with connectors |
| Microsoft | Yes (Windows) | Copilot, MCP support |
Generative Versus Designed Interfaces
The interview also touched on a design debate that matters for any AI operating system: who draws the screen.
Screens made on the fly
Startups such as Monogram and Wabi are experimenting with “generative” interfaces, which are screens an AI builds on demand rather than ones a designer fixed in advance. The appeal is obvious: an interface that fits each request exactly.
Chesky expects a mix
Chesky thinks people will see a mix of deterministic, pre-designed screens and generative ones. “There is a reason Airbnb’s app is a different design than DoorDash, Shopify, and Amazon,” he said. “Our interfaces should be adaptive to the utility and the function that we’re doing.” He does not believe apps will shrink into data layers behind one universal simple interface.
The middle ground for travel
On search, he described a destination somewhere between today’s chatbot and Airbnb’s new AI search: “I don’t want to call it a midpoint, but something between a chatbot and what you see in the first version we shipped.” Travel planning is pleasurable in itself, he noted, citing studies that people enjoy anticipating trips more than taking them, so “you don’t want to take the inspiration and the browsing and the library experience of dreaming away”.
Multiplayer AI and Voice Agents at Airbnb
Two near-term plans came out of the conversation, and both sit naturally inside an AI operating system argument because both need shared services.
Multiplayer within six months
Chesky says chatbots are built for one person, while trips are planned with family and friends. Over the next three to six months, Airbnb will explore interfaces for “multiplayer” AI that several people can use at once. That builds on the Connect and Travel Map features in the fall update, which we examined in our guide to Airbnb social features and privacy.
Voice agents this autumn
Chesky expects people to speak to computers more than they type, pointing to dictation apps such as Wispr Flow, Willow and Superwhisper. Airbnb plans to roll out voice agents this autumn for search and customer service. Spoken requests still have to be turned into structured actions, which is where natural language processing meets the booking system, and where an agent that cannot reach the app’s functions hits a wall.
Fewer meetings for the chief executive
Chesky added a personal note. “I think AI is reducing communication layers,” he said. “The way I used to get information was having meetings. Now, I can just ask my AI and I don’t need to have as many.”
How Airbnb Is Preparing for an AI Operating System
While it waits for a platform owner, Airbnb is changing its own software. Chesky said agents will need robust infrastructure on Airbnb’s side and that the company wants to be more “agent-friendly”, while also building “a lot more user interface controls”.
A model-agnostic support agent
Airbnb’s customer service agent is already the company’s clearest AI deployment. In October 2025, Chesky told Bloomberg it relied on 13 models, including OpenAI’s, Google’s and Alibaba’s Qwen, which he called “very good” and “fast and cheap”. Fortune’s account of those remarks said the agent had cut the share of users who needed a human representative by 15%.
Shipping faster with AI
The company has also seen “big gains” from using AI to ship features faster, Chesky said. Fortune reported in August that AI now writes 60% of Airbnb’s new code, a figure Chesky had given investors on an earnings call. An internal engineering team that builds with AI is better placed to expose its own functions to outside agents quickly.
What Airbnb has not said
Chesky did not announce an Airbnb MCP server, an Apps SDK integration or a date for the macro agent. The interview describes a direction, not a product, and Airbnb still prefers to bring AI into its own app rather than move its listings into someone else’s assistant.
What an AI Operating System Means for UK Businesses
Most companies will never build an AI operating system, but the argument still changes what they should prepare. Agents are arriving before the platform Chesky wants, so the practical work is making a service usable by agents without giving up the customer relationship.
Make your functions callable
List the actions a customer takes most often (check availability, book, change, cancel, pay) and expose them through a documented API or an MCP server. When an AI operating system arrives, those functions are what it will call. Our page on AI employees and autonomous agents explains how we help firms design them.
Keep the parts that sell
Chesky’s point about design applies well beyond travel. Photos, comparison tables and maps convert better than paragraphs. Plan for a handoff to your own interface at the moment of decision, rather than assuming an agent’s text reply will do the selling.
Get identity and payments right
The services Chesky listed as missing are mostly about trust: identity verification, messaging and payment. Decide now what an agent may do on a customer’s behalf, what needs explicit approval, and how you will spot a booking an agent made by mistake. Our AI strategy work starts with exactly these questions.
Watch the platform owners
The decisive AI operating system moves will come from Apple and Google. Track App Intents and AppFunctions alongside the chatbot platforms, because a phone-level AI operating system would reach far more customers than any single assistant app.
AI Operating System FAQ
What did Brian Chesky mean by an AI operating system?
He meant a platform where AI works at the core of the device, with a full developer kit so that every app can act as an agent and every agent can work with every other. Today’s assistants are apps running on iOS, Android and Windows rather than an AI operating system.
Why does Chesky say consumer AI is not working?
He has tried Airbnb inside Meta’s Muse and the Instinct agent and says it “doesn’t work very well”, worse even than hotels. He blames the missing lower layer: apps that are not agents, are not interoperable and lack rich interfaces inside agents.
What did he tell Sam Altman?
Chesky says he told Altman that an app store like the iPhone’s needs a software developer kit and an operating system, or “none of our functionality is going to exist”. Airbnb stayed out of ChatGPT’s app launch in October 2025.
Is anyone building an AI operating system?
Google’s AppFunctions lets Android 16 apps expose functions to agents such as Gemini, in private preview with testers as of May 2026. Apple’s App Intents gives Siri similar access. Neither yet matches the agent-to-agent platform Chesky describes.
Will Airbnb build its own agent?
Chesky expects a “macro Airbnb agent” that brings together the explore, customer service and other agents, interoperable with other agents through MCP. He gave no date.
Is Airbnb’s AI search available in the UK?
Not yet. The AI search features in the fall update launched for guests in the United States first.
References and Further Reading
Brian Chesky interview: AI agents need their own operating system (TechCrunch)
Airbnb 2026 fall update (Airbnb newsroom)
Chesky says OpenAI tools not ready for ChatGPT tie-up with Airbnb app (Fortune)
Overview of AppFunctions (Android Developers)
Google details AppFunctions that let Gemini use Android apps (9to5Google)
App Intents documentation (Apple Developer)
Introducing the Model Context Protocol (Anthropic)
Airbnb CEO Brian Chesky on founder mode and AI-written code (Fortune)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
