Future of the Internet: AI Agents and the Web of 2035

The Internet Is Not Disappearing. The Way We Use It Is.

From ARPANET and the dot-com crash to AI agents, digital borders and a web that may no longer need a browser.

From a small network connecting research computers to a global infrastructure serving billions, the internet has transformed how humanity communicates. Now AI agents may change how we experience it once again.

Imagine this: you need to replace a broken washing machine. You tell an assistant your budget, the size of your apartment and the delivery date. It compares independent reviews, checks which models actually fit through your doorway, finds a reputable retailer and returns with three sensible choices. You approve one. The machine arrives on Tuesday.

You never visit a search results page. You never open ten browser tabs. You may never see the retailer’s website at all.

For most of the web’s history, this would have sounded like a triumph. The internet was built to move information between computers, and the web made that information accessible to people. Now we are approaching a stranger possibility: the same network may increasingly be used by software acting on our behalf. The internet does not vanish. Its most familiar interface — the page we read, click and navigate — becomes optional for some tasks.

That sounds like progress. It also raises a difficult question: if machines do the browsing, who keeps the open web worth browsing? The future of the internet is not merely a contest between browsers and chatbots. It is a question of ownership, incentives and who gets to choose. To understand what might happen next, we have to remember that this network has reinvented itself before — and that one reinvention ended in a spectacular crash.

Before the web, there was a problem: how do computers talk?

The internet did not begin as a giant library of websites. In the 1960s, computers were expensive, isolated machines. Connecting them meant dealing with incompatible equipment, unreliable links and networks that belonged to different organizations. Researchers in several countries were exploring packet switching: instead of reserving one continuous telephone circuit for a conversation, break a message into smaller pieces that can travel across a shared network.

In October 1969, the first ARPANET message travelled between the University of California, Los Angeles, and the Stanford Research Institute. The intended word was “LOGIN.” The system crashed after the first two letters. “LO” is a wonderfully modest beginning for a technology that would later carry much of human civilization.

ARPANET mattered, but it was not the whole story. Packet-network research in Britain and France, the growth of email, satellite and radio experiments, and the work of many engineers contributed to what became the internet. The crucial breakthrough was not one better network. It was a way for different networks to communicate without becoming the same network.

In the 1970s, Vint Cerf and Robert Kahn helped develop the ideas behind TCP/IP. In January 1983, ARPANET adopted those protocols. Think of IP as the addressing and delivery system: it helps packets find their way toward a destination. TCP, where it is used, helps make sure the pieces arrive reliably and in order. Other applications can use different transport mechanisms. That separation — common rules without one owner dictating every connection — is part of the internet’s extraordinary durability.

Even its name is a clue. An internet is literally a network of networks. Nobody needed to invent a single machine large enough to contain it. They needed a system capable of cooperation across independent machines.

The web was a brilliant addition — not the internet itself

By the late 1980s, people could use the internet for email, file transfers and other services. But finding and connecting information remained awkward. In 1989, working at CERN, Tim Berners-Lee proposed linking documents through hypertext. By the end of 1990 he had built a browser-editor, a server and the basic elements of what became the World Wide Web. In 1993, CERN made the web technology freely available, helping it spread without licensing barriers.

The distinction is easy to miss because the web became so dominant. The internet is the underlying system of connected networks; the World Wide Web is one service that runs on it, alongside email, messaging, online games and many others. Websites are built from many cooperating technologies, but the simplest way to picture the relationship is this: the internet carries the traffic; the web gives people interconnected pages to explore.

Imagine typing a news site’s name into your browser. DNS works a little like a contact list, helping translate that name into the numerical IP address of a server. Your device sends requests through a home or mobile network and an internet provider. Routers pass the resulting packets between networks, often along fiber-optic cables. The page may come from the original server or from a nearby copy stored by a content delivery network (CDN), which helps it load faster. HTTPS usually encrypts the connection, while transport systems such as TCP or QUIC help deliver data reliably. Finally, the browser assembles text, images and code into the page you see. One click conceals many separate systems working together.

The physical picture is less ethereal than the word “cloud” suggests. International connections rely heavily on fiber-optic cables laid across land and beneath oceans. Data centers, mobile towers and local networks do the less glamorous work of keeping services reachable. Satellites extend access where ground networks struggle; they do not make terrestrial infrastructure obsolete. Even the cleverest internet protocol cannot route around a neighborhood without electricity.

Here was the web’s genius: anyone with a compatible server and a domain could publish a page that linked to someone else’s page. You did not need permission from one central magazine editor. A hyperlink was small, almost trivial — and socially revolutionary. It made discovery an open-ended process rather than a menu controlled entirely by one company.

So who owns the internet? Nobody owns the whole thing, but that does not mean nobody owns its parts. Telecommunications firms operate cables and mobile networks; companies and universities run servers; governments regulate access; and organizations such as the IETF and W3C develop shared technical rules. The system stays global only because many independent actors agree to communicate. Its openness is an achievement that has to be maintained, not a law of nature.

Every website visit depends on a complex network of connections. DNS helps locate the destination, routers direct data across networks, and fiber-optic cables carry information across continents. CDNs can bring content closer to users, while satellite links offer alternative connectivity.

When investors thought the future had already arrived

By the second half of the 1990s, browsers were mainstream, commercial websites were multiplying and investors could see that the network might transform shopping, media and business. They were broadly right. What they often got wrong was the speed of the transformation — and the price they were willing to pay for it.

“Put it online” became a business strategy in its own right. Startups raised money on promises of future users and future market share while the difficult question of profit was postponed. Some were selling a genuinely useful service too early. Others had a familiar old problem dressed in fashionable new language: the costs of acquiring customers exceeded what those customers were likely to spend.

The Nasdaq Composite peaked on March 10, 2000. By October 2002 it had lost roughly 78% of its value from that peak. Companies disappeared, investors lost money, employees lost jobs and the excitement about the “new economy” suddenly looked reckless. The dot-com crash was not a verdict that the internet had failed. It was a verdict that being connected to a real technological revolution did not make every valuation or business model sensible.

There is an important irony here. During the boom, telecommunications firms built networks and laid fiber in anticipation of demand that had not yet arrived. Some capacity was excessive, and the economics of that infrastructure were far more complicated than the simple “all the fiber was wasted” story suggests. But the networks, skills and software developed through the boom also helped create the conditions for the internet that followed.

Amazon survived. Google’s search business grew. Online payments became ordinary. Broadband connections improved, and later smartphones turned intermittent internet use into something close to permanent presence. The market could be wildly wrong about the short term and still be responding to an authentic long-term change.

That distinction matters in 2026, because another technology boom is reshaping the same network. AI firms and infrastructure suppliers are investing enormous sums in models, chips, data centers and electricity. In October, the International Monetary Fund highlighted the risk of a financial-market correction if optimistic expectations for AI-related earnings and investment disappoint. That is not proof that an AI crash is imminent. It is a reminder that a useful invention and a good investment are different claims.

History does not repeat itself on schedule. Today’s major AI infrastructure investors include large firms with established revenue streams, unlike many late-1990s startups. Yet the enduring question is identical: what can this technology really do for people, and can the business providing it earn more than the service costs?

The dot-com bubble peaked in March 2000 before a devastating market collapse. Companies disappeared, investment evaporated, and expectations were rewritten. Yet the underlying internet continued to expand, laying the foundations for today's digital economy.

The internet today: enormous, indispensable — and surprisingly uneven

According to the International Telecommunication Union, around 6 billion people — roughly 74% of humanity — used the internet in 2025. That is an extraordinary achievement. It also means approximately 2.2 billion people were still offline. Many others had access that was too expensive, too slow or too unreliable to support the opportunities promised by the word “connected.”

The digital divide is no longer explained only by whether a signal reaches a village. It is also about affordability, devices, language, skills and whether a connection stays alive during a crisis. Satellite-to-phone systems and new terrestrial networks may close some geographical gaps. In 2026, the ITU described direct-to-device satellite connectivity as part of a possible future convergence of space and ground networks. But coverage is not the same thing as affordable broadband, and a network that works in emergencies still needs spectrum, power, capacity and sustainable economics.

Meanwhile, the web we actually use has become much less open in its everyday experience. Search engines organize discovery. Social feeds decide what receives attention. App stores mediate software. A few cloud platforms host an immense share of services. None of those systems is the internet itself, yet each can become a gateway through which people encounter it.

The paradox is that a network designed to connect independent systems can still produce highly concentrated markets. Open technical standards do not automatically prevent economic centralization. Most people do not choose a social network because they admire its protocol. They choose it because their friends, customers or favorite creators are already there.

Freedom is another uneven layer. Access Now documented 313 internet shutdowns across 52 countries during 2025. Cloudflare’s October 2026 review also showed that power failures, natural disasters and deliberate restrictions repeatedly disrupted connectivity in the third quarter. The internet is resilient as a design principle; individual communities are not immune to a cut cable, a failed grid or a political order.

The next shift: from finding pages to delegating tasks

For thirty years, the ordinary web interaction has been remarkably stable: you have a question, you find a page, you read it and you decide what to do next. Search improved the first step. Social media changed who recommended the page. AI is beginning to change the sequence itself.

Instead of asking “Which page has the answer?”, a user can ask “Can you complete this job?” The assistant might consult several sources, compare options, fill a form, contact a service or prepare a purchase for approval. That is the practical meaning of an AI agent in this context: software with some ability to use tools and pursue a goal across steps, within permissions given by its user.

This is not merely a futuristic product demo. In September 2026, Cloudflare reported that daily requests made by AI agents across its network had risen by more than 1,700% over the previous year. It also observed that, for the first time, more than half of requests reaching sites on its network were not initiated directly by people. Those figures describe Cloudflare’s traffic and definitions, not every request across the entire internet. And much of the automated web still consists of traditional crawlers and bots, not autonomous AI assistants.

The human web is not simply disappearing. Sensor Tower’s 2026 State of Web report found that traffic to AI assistants grew 86% year over year in 2025, while visits to search and social platforms still dominated online discovery. Referrals from emerging AI sources represented less than 1% of overall web traffic in its analysis. There is no contradiction here: requests made by bots, visits to AI services and referrals from those services measure different things. The more plausible near-term future is a new layer appearing over the existing web, not an overnight replacement.

We explored how this changes discovery in AI Search: From PageRank to Autonomous Agents. The deeper issue is not whether a chatbot can summarize a website. It is whether the network’s economic and technical rules still make sense when a visitor may be an agent acting for a person who never sees the page.

Who pays for an open web when nobody clicks?

Consider a recipe writer who spends hours testing a dish. Or a local newspaper that pays a reporter to attend a public meeting. Traditionally, search engines could index their work and send them readers. Some readers saw advertising, subscribed or supported the publisher. The arrangement was imperfect, but a useful page could exchange access for attention.

An AI answer changes that bargain. The assistant may use information from multiple pages and deliver a useful answer without sending the reader to any of them. For a user, saving time is a real benefit. For the person producing the underlying information, lost traffic can mean fewer resources to create the next article. The problem becomes sharper when automated requests consume bandwidth and computing resources but generate no corresponding revenue.

In a Pew Research Center analysis of browsing activity from March 2025, Americans who encountered a Google AI Overview clicked a traditional search result in 8% of those searches, compared with 15% when no AI Overview appeared. That is an observed difference, not proof that the overview alone caused every lost click. It is enough, however, to explain why publishers are asking uncomfortable questions.

Several answers are being tested: paid access for crawlers, licensing agreements, subscriptions, services that compensate the originating publisher, and better ways for agents to identify themselves and use information with permission. None is yet an accepted economic settlement for the entire web. Micropayments sound elegant until someone must decide who gets paid for a fact repeated across a thousand sources.

There is also a cultural cost that a payment protocol cannot solve by itself. A person who receives only a synthesized answer may lose the experience of encountering an unfamiliar voice, an unexpected argument or a source that contradicts the question. We built the web to connect documents; it became valuable because documents also connected people. Convenience should not quietly remove the possibility of exploration.

The agentic web is becoming a standards problem

The original web succeeded partly because its basic language was shared. A browser did not need a private agreement with every server in order to display a page. If agents become important visitors, they need a similarly trustworthy way to understand what a service offers, what actions they are allowed to take and whether the human behind the request has actually authorized those actions.

In August 2026, Cloudflare described a possible “agentic internet” organized around four capabilities: being readable, discoverable, callable and payable by software. The proposal is not a finished universal standard. It is an attempt to define what needs to exist if automated visitors and websites are to cooperate instead of simply scraping one another.

The practical difference is easy to imagine. Today an agent might inspect a shopping page, identify an “Add to cart” button and try to reproduce a human click. A more structured system could tell an authorized agent what products are available, what conditions apply and which safe operations the merchant supports. Google’s Universal Commerce Protocol, launched in 2026, is one example of companies trying to standardize parts of that transaction. Competing approaches also exist, and consumer adoption of purchasing through AI agents has remained limited. An announcement is not the same as an everyday habit.

W3C’s October 2026 discussion of the web in the AI era identifies a similar challenge: consent, agent identity, permissions, interoperability, sustainable publishing and security. These are proposals and directions for standards work, not an announcement that the web already has a settled solution. The distinction is especially important for systems that may spend money, disclose private data or make decisions with real consequences.

A useful assistant should not require the user to understand every protocol. But it should make the boundaries of delegation understandable. “Research the options” is not the same instruction as “buy the first result.” “Draft a complaint” is not permission to send it. An agent that needs access to a calendar should not silently inherit control over a bank account.

The invisible browser creates a very visible trust problem

A web page was designed to be read, not obeyed. Imagine a travel assistant examining a hotel listing that contains a malicious instruction telling it to send the traveler’s itinerary to an unknown address. To a person, that is untrusted text on somebody else’s page. To poorly protected software, it could look like a command. Security researchers call this prompt injection. The website has no special powers; the danger arises when an agent mistakes outside information for authority and misuses permissions it already has.

This makes the future web depend on something less glamorous than intelligence: clear authority. Agents need limited credentials, isolation between sites, records of what they did and deliberate approval before irreversible actions. A system that can cancel a reservation, sign up for a subscription or send confidential files has a different risk profile from one that summarizes an article.

Trust will also involve provenance. Digitally signed credentials and content records can help establish who issued a claim or whether a file was altered after signing. The W3C has been developing tools for digital credentials, including a working draft published on October 29, 2026. But proving that a document came from a particular source does not prove that every statement inside it is true. Verification must not become a decorative seal of certainty.

The browser, strange as this may sound, was historically a kind of boundary. It let us see where we had arrived, inspect information and choose what to click. If agents act behind that visible boundary, we need a replacement for its most important function: the ability to understand what is happening and say no.

Today, we typically search, open websites, compare information and make decisions ourselves. AI agents could increasingly handle the research and comparison, presenting options for human approval. The question is not only whether this saves time, but also what happens to the websites providing the information.

When the internet can manufacture its own evidence

There is another change beyond agents. Producing a convincing article, product review, voice recording or video is becoming cheaper and easier. False information is hardly new, but the ability to make it plausible at enormous scale changes the economics of deception. A web overflowing with synthetic material could become harder to search precisely because it contains more pages, not fewer.

Digital signatures and provenance records may help establish where a file came from or whether it was changed after signing. They cannot tell us whether a photograph was misleadingly framed, whether a witness is honest or whether an argument is sound. Nor can an AI-generated label tell us whether every unlabeled item is authentic. The practical defenses remain layered: traceable origins, independent reporting, correction histories and interfaces that make it easy to compare evidence.

This is also an opportunity. Better translation and accessible interfaces could help people participate in online spaces that once excluded them by language or disability. But if the same few systems write, summarize, recommend and rank everything, diverse voices may become harder to find. An abundant internet is not necessarily an informed one. The question is whether new tools help us encounter more of the world — or merely generate a more convenient version of it.

A more connected world — or several separate internets?

Technical openness is only one force shaping the next decade. Governments are asserting more control over data flows, online speech, security and digital infrastructure. Companies have commercial incentives to keep users inside their own services. Countries may prefer domestic cloud systems and locally governed data. Some of these choices respond to legitimate privacy and security concerns; others can become tools of censorship or exclusion.

This is the “splinternet” problem: the risk that a network which can technically connect the world becomes harder to use across political, economic or platform borders. The future may not involve cables being literally cut between continents. More often, it may involve incompatible rules, closed access, regional restrictions or services that only work properly within one ecosystem.

At the same time, serious counter-movements exist. Federated services and open protocols attempt to let people communicate across independently operated platforms. Tim Berners-Lee’s Solid project proposes personal data stores that can be accessed by applications under the user’s control. Neither approach has magically solved adoption, moderation or convenience. Their importance is that they offer a different answer to the question “Who owns the relationship?”

An internet can be globally reachable and still leave users dependent on a few gatekeepers. The more capable agents become, the more consequential that dependence is. The company that controls your assistant could influence which sellers it considers, which sources it trusts, which services can connect and how transactions are completed. An agent that works for you should not quietly become a salesperson for the platform that supplies it.

The physical internet will change too

It is tempting to imagine the future internet as a software story alone. But someone must manufacture the chips, run the servers, lay the fiber, manage the spectrum and supply the electricity. New interfaces do not repeal physics.

Connectivity will keep expanding through a combination of fiber, mobile networks and satellites. Direct-to-device communications could make ordinary phones more useful beyond traditional cellular coverage, especially for messaging and emergency access. Higher-capacity satellite systems may serve remote communities, ships, aircraft and infrastructure. The useful future is not “everything moves to space.” It is a more resilient combination of ground and space networks, with each doing the jobs it does best. Next Horizon’s Internet of Things explainer examines how the same network is reaching machines and sensors that people never browse directly.

Underneath that growth, the encryption used to protect connections is changing too. In August 2026, the IETF published RFC 10024, describing ways for TLS 1.3 — a technology that protects many secure internet connections — to combine established cryptography with newer, quantum-resistant methods. The goal is to prepare for a future in which powerful quantum computers might threaten some of today’s widely used encryption. This is an upgrade to the security of the existing internet, not evidence that a completely separate “quantum internet” is about to replace fiber and Wi-Fi.

The energy demands will be substantial. In its 2026 update, the International Energy Agency projected that data-center electricity use could rise from about 485 terawatt-hours in 2025 to roughly 950 terawatt-hours by 2030 — approximately 3% of global electricity demand in that year. AI is an important driver, but data centers also support the rest of the digital economy. This is a central projection, not a meter reading from the future. Efficiency may make each task cheaper while overall electricity demand still rises as more tasks are performed.

The question is not whether the internet should use electricity — every useful civilization-scale system does. It is whether the value created justifies the infrastructure, and whether that infrastructure is affordable, secure and sustainable. A future that saves a person five minutes but requires extravagant compute for every routine request may need better engineering, not better marketing.

Three plausible futures for the 2030s

The easiest way to get the future wrong is to choose one impressive demo and draw a straight line from it. Internet history suggests a more cautious approach. A technology can arrive quickly, develop slowly in the details and then change society in ways its inventors never intended.

1. The open, assisted web

In the most constructive scenario, people use agents for repetitive digital work but retain meaningful choice over providers, data and sources. Open standards let a travel agent compare services without requiring exclusive deals. Publishers can state how their material may be accessed and receive fair compensation when reuse has commercial value. Websites remain important for original reporting, expertise, culture, public information and experiences worth having directly. The agent is a useful tool at the edge of the web, not its owner.

2. The platform-controlled web

In a less appealing scenario, a small number of assistants become the default gateway to everything. Their operators negotiate with a limited group of businesses and information suppliers. Convenience improves, but visibility declines: users see fewer independent options, creators become dependent on opaque ranking rules, and the price of reaching customers shifts from search advertising to paying for access to the agent. The internet still works technically. Economically, much of its freedom is filtered through private interfaces.

3. The fragmented, defensive web

A third possibility is an increasingly defensive network. Publishers restrict automated access. Governments impose incompatible identity and content requirements. Services become harder to use across borders. Security warnings and permission prompts proliferate because neither people nor businesses fully trust automated visitors. The result is more friction, not necessarily less intelligence.

These outcomes are not mutually exclusive. The internet of the 2030s may combine open scientific repositories, tightly controlled financial services, national restrictions, personal AI assistants and thriving human communities. Different parts of the network will evolve at different speeds. The important decisions are being made now, in the standards, incentives and product designs that determine what can connect to what.

What should we actually watch over the next five years?

I would watch a few unglamorous indicators rather than a hundred dazzling product launches. Can an agent finish a task across several independent services without special deals? Can people change assistants without abandoning their data, preferences and permissions? Can publishers benefit financially when automated systems rely on their work? Can users inspect the original sources and approve purchases before money moves? And can another billion people get a useful, affordable connection — not merely a signal on a coverage map?

Those are better tests than whether a demo looks like science fiction. Internet adoption was not driven solely by the day a computer could show a web page. It took years of better interfaces, cheaper connections, useful businesses and ordinary people finding reasons to come back. Agents will face the same test. Their success should be measured not by how many steps they can hide, but by how much practical control and value they give the person who asked for help.

By 2035, the internet could become more intelligent, accessible and interconnected. But it could also grow more centralized, with a handful of platforms controlling how people and AI agents discover information. The technology will matter. So will the choices we make about openness, competition and digital freedom.

The internet’s future is still a human decision

The early internet had no single master plan for shopping, music, journalism, video calls or online communities. It created a common set of rules, and people built unexpectedly useful things on top. The web added a simple and radical idea: information should link to information, and anyone should be able to participate.

Now we face the possibility of a network where some of those participants are machines acting on our instructions. I find that genuinely exciting. A good assistant could make unfamiliar systems easier to navigate, help people compare choices and open services to those who struggle with today’s interfaces. But those gains are not automatic. Removing clicks is not the same thing as removing barriers. And an internet that saves us from visiting websites is not necessarily an internet that gives us more freedom.

The dot-com crash left us a useful lesson: technology can be transformative even when predictions about its immediate future are wildly overpriced. The next lesson may be political and cultural as much as financial. We will have to decide whether intelligent intermediaries expand our ability to explore the world or quietly decide which part of it we are allowed to see.

The future internet will still be made of cables, computers, protocols and people. The new question is who gets to act, who gets to choose — and whether the extraordinary openness that made the web valuable survives its next reinvention.