The Silent Build of the Agentic Web and the Coming Shift in Travel Distribution

TL;DR

• The agentic web isn’t stalled — it’s being built quietly beneath the consumer surface. Protocols like MCP (97M+ monthly SDK downloads, adopted by every major AI platform in under a year), Google’s A2A for agent-to-agent coordination, and payment infrastructure like AP2 are filling in fast. Most builders will never make the headlines.

• 56% of U.S. travelers already use AI for some part of their travel journey, up from 33% a year ago. The operating model looks like metasearch: agents handle discovery, research, and recommendations, then hand the traveler off to book. The shopping is the disruption, not the checkout button.

• Power is shifting from whoever controls the interface (OTAs, search engines) to whoever controls the consumer-side agent and its harness — the tools, rules, policies, and data that determine what the agent recommends. These are the same levers OTAs used to dominate hotel distribution. They’re just moving to a new address.

• Travel has no durable human-only cognitive moat. AI is cracking protein folding and compressing drug trials. Hotel distribution is arcane rules and siloed data — complicated like tax codes, not hard like molecular biology. Once the infrastructure is in place, agents will match or beat human specialists.

• OTAs have a near-term advantage because they’ve already built agent-accessible infrastructure (MCP connectors, structured content, ChatGPT integrations). Suppliers that haven’t started are already behind in the agentic channel.

• Supplier-side investments in structured data, offer management, and personalization pay off twice: they improve the human web now and plug directly into agentic channels the moment the infrastructure is ready. The CFO doesn’t have to bet on the future — they can invest in something that works today and positions them for what’s next.

Many people I talk to in hospitality right now think AI agents are stalled. They’ve seen the demos, tried a few, gotten a mediocre hotel recommendation from a chatbot, and moved on. “Agents aren’t there yet” has become a common position in the industry.

They’re right about the current consumer experience. And they’re drawing exactly the wrong conclusion from it.

What they’re missing is happening below the surface. Not in the chatbot window, but in the infrastructure layer — the protocols, registries, governance systems, and orchestration patterns that agents need before they can do anything useful at scale. That infrastructure is being assembled right now, at speed, we’re just not seeing it because it isn’t being delivered with the fanfare and fireworks of a new model.

JFrog’s launch of a Universal MCP Registry on March 18 is one recent example. MCP (the Model Context Protocol) is the standard that lets AI models connect to outside tools and data. Think of it as the plumbing that allows an agent to reach into a hotel’s inventory system, a payment network, or a review database. JFrog is now treating those connectors with the same security controls as enterprise software packages — providing governance, discovery, and trust controls so companies can adopt agent tooling without flying blind (JFrog, March 18, 2026). That’s not a consumer product announcement. It’s plumbing. And it matters enormously because plumbing determines whether anything above it can work.

The calm surface is misleading. Remember the movie Jaws? Yeah, it’s like that.

One Internet, Two Interaction Layers

I want to be precise about what’s happening, because it can be a little confusing.

The internet is not splitting into two networks. There is one internet: the same underlying rails, the same cloud infrastructure, the same transport layer. What’s emerging is a second interaction layer on top of that shared (internet) foundation.

The human web is the layer we’ve lived with for decades: pages, apps, feeds, search boxes, menus, ads, click-driven journeys. It was built around human attention. This is the web where value goes to whoever captures eyeballs and converts them into clicks.

The agentic web (actually a new interaction layer) is something different. It’s built around machine-to-machine interaction: APIs, standardized connectors, registries, orchestration logic, etc. I’ll call it the “agentic web” for convenience, but it’s not literally a second web of pages. It’s a new way of interacting with the same underlying internet, one where software acts on behalf of a person rather than the person doing the clicking.

Both layers will coexist. The human web is not going away; it will continue to suck up cash from your budget and deposit it into Google’s coffers.

Between them sits a (somewhat) useful bridge: computer-use agents. These are agents that can operate against the human web directly — clicking through websites, filling in forms by navigating interfaces built for people. Both Anthropic and OpenAI shipped production versions of this capability in recent weeks (CNBC, March 24, 2026). Computer-use agents matter because the human web already exists at an enormous scale, and the agentic layer is still incomplete. They’re a way for agents to be useful now while the AI-native plumbing catches up.

But computer-use agents are a transitional crutch. For mainstream commerce at scale, native agent-to-system communication will outperform screen-scraping every time. Computer-use will persist for edge cases and legacy systems. For the core of travel commerce, the agentic layer is where this is headed.

What Makes the Agentic Layer Different

Here’s what matters most, at least the way I see it.

Agents don’t browse. Humans navigate interfaces, click through menus, scan pages, and make sense of visual layouts. Agents query systems, gather structured inputs, and optimize for outcomes. The entire UX toolkit that travel companies have spent decades perfecting — hero images, persuasive copy, funnel optimization, A/B-tested booking flows — is invisible to an agent. Agents don’t have eyeballs to catch.

Ads don’t work inside the agentic layer. On the human web, ads work because human attention is the scarce resource and half the time there are so many ads you can hardly tell which elements on the page are organic vs. not. Inside agentic decision flows, the inputs that matter are intent, memory, trust, permissions, routing logic, and data quality. Agents don’t care about banners, placement, or persuasive UI. They couldn’t care less how many other people have that item in their shopping basket or how many bought it so far today. Advertising may still influence people upstream — travelers will still see ads on Instagram — but that’s the human web doing what it does. Inside the agentic layer, those signals carry no weight.

Power shifts to whoever controls the agent. On the human web, power sits with whoever controls the interface: search engines, OTAs, marketplaces, apps. In the agentic layer, power shifts toward the consumer-side agent and the orchestration environment it runs inside. More on that in a minute.

One additional point worth emphasizing: identity becomes the price of admission. If an agent wants to present a request to a supplier and get something personalized in return, it has to identify who’s asking. It needs to present credentials like a loyalty program ID so the supplier’s systems can fashion a tailored offer based on what they know about that specific traveler. Show up without authenticating anyone, and you’ll get the plain vanilla response: standard rate, standard room, no personalization.

The Agentic Layer Is Being Built Now

This is the section that matters most. Time to focus!

Many people are right that agents don’t yet deliver a consistently reliable consumer experience for complex travel shopping. The wrong conclusion is that little is happening. The right one is that the agentic layer is still building the infrastructure that the human web accumulated over decades: standards, interfaces, tool connectivity, governance, trust mechanisms, and discovery layers.

Three categories of infrastructure tell the story. Not the whole story, obviously, but enough to give you a flavor for what’s happening.

MCP as a standard for connecting models to tools and data. Anthropic released MCP in November 2024. Within a year, OpenAI, Google, Microsoft, and every major AI platform adopted it. It now has over 5,800 connectors, 97 million monthly SDK downloads, and governance under the Linux Foundation through the Agentic AI Foundation — co-founded by Anthropic, Block, and OpenAI, with support from Google, Microsoft, and AWS (Anthropic, December 2025). Few standards in tech history have achieved comparable cross-vendor adoption this quickly (The New Stack, December 2025). MCP is the connective tissue that lets agents interact with outside systems. For example, if a consumer agent from, say, OpenAI, wants to get offers from a hotel system, it’s going to connect with an MCP server on the supplier side, and that MCP server will connect to systems like CRS, CMS, and RMS. A year ago, MCP barely existed; now it’s the industry standard. This insane speed of progress will be a common theme as we move forward.

A2A as a pattern for agent-to-agent coordination. If MCP lets an agent talk to a tool, A2A — the Agent2Agent protocol from Google — lets agents talk to each other. Google launched it in April 2025 with over 50 partners, and by mid-2025, it was under the Linux Foundation with 150+ supporting organizations (Google Cloud Blog, July 2025). Think of the practical implications: a traveler’s personal agent could coordinate directly with a hotel’s agent to negotiate a package, check availability against a calendar, and confirm preferences — all without a human interface in the middle. Google then layered on the Agent Payments Protocol in September 2025, with more than 60 organizations — including Mastercard, PayPal, and American Express- working out how agents securely initiate and complete payments (Google Cloud Blog, September 2025). That’s MCP for tools, A2A for agent coordination, and AP2 for money. The stack is filling in.

Registries and governance layers. The JFrog announcement I opened with is one visible example, but it represents a much bigger movement. Dozens of companies are spending real money building the governance, security, and discovery components that agent-based systems need to work reliably at enterprise scale. Most of this work will never make a headline. Gartner published guidance in November 2025 advising that security leaders should establish centralized registries for agent connectors with layered security controls. GitHub launched its own MCP registry. The official community registry grew to nearly two thousand entries within months of its September launch. These are the architectural components that will make everything work once the build reaches critical mass — and most of the people and companies building them will never get any press for doing so. How many of you thought JFrog was a new toy when you first read the name above? See?

Meanwhile, the demand signal from travelers is well past theoretical. Phocuswright’s March 2026 report — titled The AI Surge: Travel’s Fastest Behavioral Shift in a Decade — found that 56% of U.S. travelers now use AI for some part of their travel journey, up from 33% in early 2025 (Phocuswright, March 2026; PhocusWire, March 2026). Every generation posted double-digit gains. Adobe reports that generative AI traffic to U.S. travel sites surged 3,500% year-over-year by mid-2025, with 29% of consumers using AI for travel tasks and 88% of those saying it improved their experience (Adobe, 2025).

How travelers are using AI tells you where this is going. Right now, travelers use AI heavily for discovery, research, and comparison — and then go to the supplier or OTA to complete the transaction. This is the metasearch model all over again: the agent handles the cognitively heavy shopping and makes the recommendations, then delivers the customer to the point of purchase for the option they choose. For suppliers, the big issue is whether and how they are represented in the recommendations. Anyone waiting for “agents that book autonomously” before taking this seriously is like a hotelier in 1998 dismissing the web because nobody had figured out online payments yet. Those are the hoteliers who let the OTAs silently build their dominance while they waited. The shopping recommendations are the disruption. The checkout button is a detail. Over time, the model will develop toward more autonomous purchasing for travelers who are willing to delegate that authority within whatever parameters make them comfortable. The metasearch pattern is the starting point, not the end point. But I’d also acknowledge that plenty of people will want a hand in the planning and booking process regardless of how much they trust their agent, and that’s only natural. The point is that agents will handle an increasing share of the heavy lifting.

The Real Control Point: The Consumer Agent Plus the Harness

Memory gets a lot of attention in discussions about agentic commerce, and for good reason. Memory establishes continuity, preference patterns, tradeoff history, and context. I wrote about memory’s role in reshaping travel distribution in a previous article. But memory alone is not the control point.

The control point is the consumer-side agent operating within what’s called a harness—the full set of tools, rules, policies, data, and permissions that shape how the agent does its work.

Think of the harness as everything the agent has access to and everything that constrains it. Which supplier systems can it query? What policies govern how it ranks options? What data does it draw on to personalize recommendations? What rules determine when it asks the traveler for confirmation versus acting on its own? What files and reference material inform its understanding of the traveler’s preferences? The harness is the operating environment that turns a general-purpose model into a travel-shopping agent that works for a specific person.

You can already see this taking shape. ChatGPT now has an app ecosystem — built on MCP — that determines which third-party systems (Expedia, Booking.com, and others) the agent can access (OpenAI, October 2025). It has a memory system that accumulates traveler context across sessions. It has shopping-specific interfaces for comparing results. And OpenAI controls the policies, ranking logic, and rules that determine which sources are called and in what order. Claude, from Anthropic, has its own version: over 75 MCP connectors, a computer-use capability for operating against sites that don’t have native agent interfaces, and its own set of tool-routing and policy controls.

The party that controls the LLM and harness has disproportionate influence over routing, ranking, trust, defaults, and supplier visibility. If that sounds like a familiar set of levers, it should. These are the same levers OTAs used to dominate hotel distribution on the human web. The difference is that the levers are moving to a new location, and the parties holding them will likely be different. The question for hotel executives isn’t whether this shift is real, it’s who ends up holding your levers this time and how you can influence them.

Why Travel Is One of the Most Consequential Battlegrounds

Travel is consequential for agentic commerce because it sits at the intersection of several factors: the market is enormous (global travel gross bookings reached $1.6 trillion in 2024, per Phocuswright (Phocuswright, October 2025)), shopping is cognitively heavy, supply is deeply fragmented, comparison is messy, and personalization can materially improve outcomes for both consumers and suppliers.

There’s a bedtime story that still circulates in some executive suites: travel is too complex for AI agents. Too many edge cases, too many variables. It’s a story that lets reluctant executives sleep at night.

Travel is genuinely complex. But there is nothing about that complexity that amounts to a durable human-only cognitive moat. Once you’ve seen an LLM read a few CRS contracts, compare primary provisions, and suggest changes that will help you with your specific needs, you begin to appreciate how well AI does with complexity. (Note: there are some companies that appear (to me) to thrive on the complexity of their contracts. I’m not naming names, but we’ll have moracle to say about this later…)

Still think travel is beyond the abilities of AI to understand? Consider what AI is doing in domains that are orders of magnitude harder. Protein folding — predicting the three-dimensional structure of molecules from amino acid sequences — stumped human scientists for fifty years. AI cracked it. Drug discovery, materials science, and mathematical proof — these require genuine scientific reasoning at levels most humans can’t reach. On March 26th, The Information (paywall) reported that AI is saving drugmaker Novo Nordisk tens to hundreds of millions of dollars in conducting drug trials. They reported: “Now, Novo Nordisk says it’s using AI agents to shorten by weeks or months the time it takes to both start new clinical trials and to complete them.” That’s complex stuff. Imagine all the rules involved in drug trials! Travel, by contrast, is a series of arcane rules, fragmented data sources, and legacy distribution logic. It’s complicated in the way that tax codes are complicated: the product of decades of accumulated decisions, not some deep mystery of the universe. It is not hard in the way that protein folding is hard, nor is it as complex as drug trials.

The right combination of a frontier model, a capable agent, a well-built harness, and the right tools and data access will be able to meet or surpass human abilities in the travel domain. The agent never sleeps, never forgets what you told it three trips ago, and can run thirty supplier queries simultaneously while comparing the results against your preferences and behavioral history. The constraint right now is infrastructure. As the plumbing fills in, travel complexity won’t be a moat. It’ll just be a solvable problem.

Some executives may be mistaking infrastructure immaturity for cognitive impossibility. That’s a dangerous error because it gives them a false sense of how much time they have.

Rich Intent Is the Disruptive Force

Traditional travel shopping relies on thin signals crammed into an annoying widget: destination, dates, room count, basic filters, maybe loyalty status. That’s what search boxes were built to capture. It’s also a remarkably low-resolution picture of what a traveler actually wants.

Consumer agents can surface much richer intent: neighborhood preferences, sleep sensitivity, reason for trip, schedule constraints, family composition, cancellation sensitivity, budget flexibility, amenity tradeoffs, brand leanings, room adjacency needs, and dozens of other signals that live in memory and context but never made it into a search box.

Richer intent means better-matched options and, over time, better offers. This is the beginning of better merchandising — offers built around what this specific traveler actually values rather than what a generic filter set can capture. I remember a senior executive from an investment firm covering travel who said, “AI can do a lot of things, but it can never deliver those magic moments that make travel special.” He’s absolutely right, but those magic moments won’t occur unless the right traveler is delivered to the right destination with the right environment to produce them. And which mechanism do you think is more likely to match that traveler and that magical destination: a low-fidelity OTA travel widget or a consumer-side AI agent with a deep understanding of the traveler and their needs?

The hotel that can receive rich intent and respond with a structured, personalized offer has a material advantage over the one returning a standard rate and room type. That’s true on the human web today. It will be dramatically more true in the agentic layer.

Aggregation Shifts Into the Agent Layer

OTAs are not vanishing. Suppliers are not bypassing all intermediaries overnight. The human web remains intact. Take a breath. Let me be clear about that, because every time I write about agentic commerce, someone implies I’m writing the OTAs’ obituary. I’m not. But the aggregation function, previously dominated by intermediaries with huge ad budgets, is moving. And with it will move the corresponding revenues.

There’s an important distinction here. Access means reaching multiple sources. Aggregation means collecting, normalizing, comparing, ranking, and recommending. Access is more of a commodity. Aggregation is where the value sits.

If the consumer-side agent can call multiple sources, normalize what comes back, compare it against the traveler’s intent, and rank the options, then the aggregation function is occurring in the agent layer. The value creation that used to justify OTA commissions is happening somewhere new.

The metasearch analogy helps. Metasearch showed that discovery and comparison could sit in one place while fulfillment sat somewhere else. The agentic layer takes that further: the agent performs the discovery, comparison, and recommendation, while the transaction may still flow through a supplier-direct channel or an intermediary. For now.

What This Means for OTAs and Suppliers

For OTAs: They remain important on the human web. And in the near term, they have real advantages in the agentic layer — because they’ve done the work. OTAs have already structured their content for machine consumption and built the MCP connectors that make their inventory accessible to agents. Booking.com and Expedia are already integrated into ChatGPT’s app ecosystem. Until suppliers build the same kind of agent-accessible infrastructure, OTAs will be the path of least resistance for consumer agents looking to fulfill travel intent.

Business will shift from web channels to agentic channels over time. The point of this article is that the shift won’t happen wholesale until more of the underlying architecture is built. But the architecture is being built. And the OTAs that are prepared will capture early share, while the suppliers that aren’t will find their properties invisible to an increasingly important channel.

For suppliers: The incentive to prepare goes beyond margins and control. It includes direct relationships, first-party data advantages, richer offer construction, stronger personalization, and new customer acquisition, which OTA-mediated channels constrain. When the OTA controls the interface, the OTA controls who sees your property and under what conditions. When the traveler’s agent controls the interface, the agent’s ranking logic determines visibility.

Here’s a practical reason to move now: supplier-side investments pay off twice.

Investments in structured data, offer management systems, personalization engines, and identity-aware direct experiences improve the human web right now. If a supplier builds a better offer management system that can construct personalized packages based on traveler context, that system can also work on the supplier’s own website. The revenue uplift from better web merchandising may justify the investment on its own terms. Then, when the agentic infrastructure is ready — and it’s getting ready faster than most operators realize — those same systems become immediately available to agentic channels from day one. The CFO doesn’t have to make a bet purely on the future. They can make an investment that pays off on the web now and positions them for what comes next.

And there’s another angle. AI and agents may not just learn to navigate travel’s inherent complexity — over time, they may help suppliers remove some of it. A lot of the arcane rules and rigid structures in hotel distribution exist because older systems and distribution models forced them into place. As agent-native infrastructure improves, some of that legacy complexity may become optional. That would be a welcome development for anyone who has spent time explaining the rationale of certain (this-could-only-happen-here) length-of-stay restrictions and derived rate logic to a revenue management team at 2 AM.

The Two Layers Compared (A Quick Recap)

Gradual Build, Then Consumer Inflection

The infrastructure side moves gradually. Standards get written. Registries fill up. Governance frameworks and orchestration patterns mature. Supplier systems get upgraded. Trust and identity layers develop. This happens over months, nearly all of it out of public view.

The consumer side will feel more sudden. Consumers won’t track the infrastructure build step by step. They’ll notice when recommendations become materially better, when itinerary construction gets much easier, when comparison improves noticeably, and when agent-mediated shopping becomes reliably useful. That transition — from “meh” to “wow, that actually worked” — can happen faster than executives expect, because the groundwork is already being laid. Remember that Phocuswright and Adobe showed that many people are already deriving significant benefits from using chatbots for travel planning, but there’s still a lot more to be done.

This is the same pattern that played out with mobile commerce, with ridesharing, and with the first generation of OTAs. The infrastructure build happens gradually. The consumer adoption curve bends sharply. By the time the experience gets good enough, the underlying architecture is already locked in, and the companies that built for it early have structural advantages that latecomers will lose plenty of sleep over.

Do not misread the calm.