AI 旅遊的終局：超級應用程式還是超級智能體？我押注超級智能體

There's a debate happening in the travel tech industry right now that will determine the next decade of winners and losers. It's not about which AI model is best. It's not about which OTA has the most inventory. It's about a fundamental architectural question:

Will the future of travel booking be a Super App — a single, all-in-one application where users search, book, and manage their travel, enhanced with AI features?

Or will it be a Super Agent — an AI assistant that lives in your conversation interface, connects to any service through standardized protocols, and handles every aspect of your trip without you ever opening a dedicated app?

I'm on record betting on the Super Agent. Not because it's more exciting, but because the technology, the economics, and the user behavior all point in one direction. This is my full argument — the case for why the Super App model is a dead end and the Super Agent model is inevitable.

Defining the Two Visions

Before I argue, let me define the two visions clearly, because the terms are often used loosely.

The Super App vision: Travel booking remains centered on dedicated applications — Expedia, Booking.com, Airbnb, Google Travel. These apps add AI features: conversational search, personalized recommendations, AI-powered customer service. But the user still opens the app, still interacts with the app's interface, and still completes the booking within the app's ecosystem. The app is the destination. AI is a feature.

The Super Agent vision: Travel booking moves out of dedicated apps and into conversation interfaces — Claude, ChatGPT, Siri, Google Assistant, or specialized AI agents. The user describes what they want in natural language. The agent connects to travel services (hotels, flights, rental cars, restaurants) through standardized protocols like MCP, searches across all of them, compares options, and completes bookings — all within the conversation. The app is infrastructure. The agent is the interface.

These are not just different product designs. They're different business models, different power structures, and different futures for who controls the $8 trillion global travel industry.

Why the Super App Model Is Doomed

Let me make the case against the Super App first, because it's the incumbent model and the one most people assume will win.

Reason 1: Users don't want more apps. The average smartphone user has 40+ apps installed but uses only 8–10 regularly. Travel apps are among the least frequently opened — most people book travel 2–4 times per year. Asking users to download, learn, and regularly open a dedicated travel app is fighting against fundamental user behavior. Users consolidate. They don't proliferate.

Reason 2: AI makes app interfaces obsolete. The entire value proposition of a travel app UI — search forms, filter panels, comparison tables, booking flows — is designed for humans to navigate manually. An AI agent doesn't need a UI. It can parse a natural language request, query 500 hotels, compare them across 15 dimensions, and present the top 3 — all in seconds. The app interface is overhead for an agent. It's like having a steering wheel in a self-driving car.

Reason 3: No single app can have everything. Even the biggest Super App — whether it's Expedia, Booking, or a future WeChat-style mega-app — can't integrate every travel service. There are millions of hotels, thousands of airlines, hundreds of thousands of tour operators, and an endless long tail of local experiences. A Super App can integrate the top 10% and ignore the rest. A Super Agent connected through MCP can access everything, because MCP is an open protocol that any provider can join.

Reason 4: The economics favor agents. A Super App spends billions on customer acquisition, marketing, and app store optimization. A Super Agent acquires users through the conversation platforms they already use (ChatGPT, Claude, Siri) at near-zero marginal cost. The Super App's CAC is $30–$80 per user. The Super Agent's CAC is effectively zero, because the user is already in the conversation interface.

Reason 5: Control is shifting to the user. In the Super App model, the app controls what you see, in what order, at what price. Sponsored placements, algorithmic ranking, and cross-sell funnels are designed to maximize the app's revenue, not your experience. In the Super Agent model, the agent works for you — it searches across all providers, finds the best price, and has no incentive to promote one supplier over another (unless you tell it to). Users are increasingly aware of this difference, and they're choosing agents.

Why the Super Agent Model Is Inevitable

Now the positive case. The Super Agent isn't just better than the Super App — it's the natural endpoint of several converging trends.

Trend 1: Conversation is becoming the universal interface. We're already seeing this with ChatGPT, Claude, and Siri. Users increasingly prefer to type or speak what they want rather than navigate menus and forms. This is especially true for complex tasks like travel planning, where a single conversation can replace dozens of app interactions.

Trend 2: MCP standardizes service connectivity. The Model Context Protocol does for AI agents what HTTP did for web browsers — it creates a standard way for any agent to connect to any service. Before MCP, an AI agent needed custom integration code for every hotel, airline, and restaurant. After MCP, it needs one configuration block. This is the infrastructure that makes Super Agents possible.

Trend 3: Agents are getting better at execution. Early AI agents could only search and recommend. Modern agents can execute — book hotels, reserve restaurants, buy tickets, process payments. As execution capability improves, the gap between "agent can plan your trip" and "agent can book your trip" closes. And once an agent can both plan and book, the Super App becomes redundant.

Trend 4: Infrastructure is being unbundled. The travel industry is moving from vertically integrated platforms (OTA = search + booking + payment + customer service) to horizontally layered infrastructure (MCP server = inventory, payment provider = payment, AI agent = interface). This unbundling makes Super Agents possible, because each layer can be best-in-class independently, rather than being locked into one platform's mediocre everything.

Trend 5: Younger users expect conversational interfaces. Users under 30 have grown up with messaging apps, voice assistants, and chatbots. They expect to interact with services through conversation, not forms. For this demographic, opening a travel app and filling out a search form feels as antiquated as sending a fax. They will use agents. They will not use Super Apps.

The Timeline: When This Happens

I don't believe in overnight revolutions. The transition from Super App to Super Agent will be gradual, but the direction is clear:

2026 (now): Super Apps add AI features. Super Agents can search and recommend but struggle with complex bookings. Most travelers still use apps. MCP adoption is growing but not yet mainstream.

2027: Super Agents handle routine bookings (single hotel, simple flights) reliably. MCP servers cover major travel verticals. 15–20% of online travel bookings involve an AI agent at some point in the funnel. Super Apps begin offering MCP-compatible APIs as a defensive measure.

2028: Super Agents handle complex multi-service bookings (flight + hotel + rental car + activities). MCP is the standard protocol for AI-travel service connectivity. 30–40% of online travel bookings are agent-mediated. Super App traffic declines as users shift to conversation interfaces.

2029–2030: Super Agents are the primary booking interface for a majority of online travelers. Super Apps become infrastructure providers — their inventory and booking engines power agents, but consumers rarely open the apps directly. The travel industry's power shifts from app owners to agent platforms and MCP infrastructure providers.

This timeline is aggressive but realistic. The technology exists today. The adoption curve is the only variable, and adoption curves for conversational interfaces have been faster than anyone predicted.

What This Means for Different Players

For OTAs (Expedia, Booking, Trip.com): The Super Agent transition is an existential threat. Their entire business model is built on controlling the booking interface. If the interface moves to agents, they become commodity inventory providers. Their options: (1) build their own agents and hope users prefer them, (2) become MCP infrastructure providers and power other agents, or (3) get marginalized. Most will try (1) and fail, then reluctantly do (2).

For hotels and airlines: The Super Agent transition is an opportunity. Direct connections through MCP let hotels and airlines reach travelers without paying OTA commissions. A hotel that connects its inventory to an MCP server can be booked by any agent, anywhere, at a fraction of the OTA cost. This is the biggest shift in distribution power since OTAs displaced travel agencies.

For AI agent platforms (OpenAI, Anthropic, Google): The Super Agent transition makes them the new gatekeepers. If users book travel through ChatGPT or Claude, these platforms control the interface and can monetize through transaction fees, premium agent features, or advertising. They become the new OTAs — but without the inventory cost.

For MCP infrastructure providers: This is the sweet spot. Every Super Agent needs inventory, booking, payment, and customer service infrastructure. MCP servers that provide these capabilities become the foundational layer of the travel industry. The companies that own this infrastructure will be the most profitable players in the ecosystem.

For travelers: The Super Agent transition is unambiguously good. Lower prices (no OTA markup), better recommendations (agents search across all providers), more convenience (conversation instead of forms), and more control (the agent works for you, not for the platform).

The Counterargument (and Why It's Wrong)

The strongest counterargument for the Super App model is trust. "Users trust Expedia to handle their booking. They won't trust an AI agent to spend their money."

This was true in 2024. It's less true in 2026. It will be false by 2028. Here's why:

Trust in AI agents is improving rapidly. As agents get better at executing bookings correctly, as MCP infrastructure provides verified inventory and confirmed pricing, and as users have more positive experiences, trust builds. The same thing happened with online banking (people initially didn't trust it), e-commerce (people initially didn't trust entering credit cards online), and ride-sharing (people initially didn't trust getting into a stranger's car). Trust follows reliability, and reliability is improving fast.

Moreover, Super Apps aren't inherently trustworthy. They have their own problems — hidden fees, dynamic pricing, sponsored placements that aren't labeled, customer service that's hard to reach. The trust advantage of Super Apps is real but eroding, and the convenience advantage of Super Agents is growing. At some point, the lines cross.

The Bottom Line

I'm betting on the Super Agent because I believe the interface layer of travel will move from apps to conversations, and the infrastructure layer will move from proprietary platforms to open protocols. This isn't a guess — it's the pattern of every major technology shift. Mainframes → PCs. Web portals → search engines. Native apps → conversation interfaces. Each shift moves power from closed platforms to open infrastructure.

The Super App model is the last gasp of the closed-platform era. It's adding AI features to a fundamentally outdated architecture. The Super Agent model is the beginning of the open-infrastructure era — agents that connect to any service, through any protocol, working for the user rather than the platform.

The hotels will still be there. The flights will still fly. The money will still be spent. The only question is who controls the interface through which it all happens.

I know where I'm placing my bet.

The MCP infrastructure that powers Super Agent travel booking — real-time hotel inventory, confirmed pricing, and transaction capability — is available as open-source software: RollingGo Hotel MCP on GitHub.