The web was built for people. A tsunami of AI agents is changing that fast.

I visited the CEO of a major startup in SF last week. He recounted a tale that blew my mind.

His company recently put a new product online but hadn’t done a big launch or any marketing. Within a week or two, it got at least a million signups. Barely any humans even knew about it. This was all AI agents just finding this new tool and using it en masse.

This is the future of the internet, and it’s arriving at warp speed.

Back in early 2025, I described a vision of a “dark internet” where humans will be drowned out by agents accessing and using online information, widgets, and tools all by themselves.

Silicon Valley got busy building the “rails” for this agentic web so that these autonomous software programs could act on our behalf, executing workflows, doing online chores, and even paying for things, all without us lifting a finger.

This took a couple of years, but the early foundations have been built and that’s been enough to support a sudden surge in agent activity online, producing both stunning and worrying results.

Meta recently launched a personal agent called Muse, which got busy scouring the web and calling customer service on behalf of millions of users. It hit the top of the app charts late Thursday. Instinct, an agent that lives in your text messages, is performing similar feats.

In the workplace, coding agents are churning out millions of lines of code faster than any human software engineer. And some of these bots have broken out of online testing environments, sparking new predictions of AI doom and calls for an industry slowdown.

‘Rounding error’

That’s unlikely to stop the agentic internet from exploding and changing how our lives and businesses are run in unprecedented ways.

Bot traffic overtook human traffic online a few months ago, and it could become 1,000 times bigger than human-generated traffic in about five years, according to Cloudflare estimates.

“Humans will be a rounding error on the internet, not because human traffic goes down, but that’s just how fast we’re seeing non-human traffic grow,” Cloudflare CFO Thomas Seifert said earlier this year.

“AI agentic Internet traffic will obviously VASTLY exceed human usage. Not a close call at all,” Elon Musk wrote on X following the Cloudflare projection.

An internet for machines

An internet designed for machines is likely to be very different from the one that was built over the past few decades for humans.

Take a look at this video posted this week on X by Gregor Žunič, cofounder of Browser Use, a startup that’s developing an AI agent framework for controlling web browsers and automating online tasks.

It shows an agent using Google to find the best flight from Zurich to London on a specific date. It took the bot seven seconds, way faster than a human could do it.

Imagine this happening millions of times a second on sites built for human, not machine, speed?

Muse saves

As with all powerful new technology, the agentic internet will produce good and bad outcomes. And we’re already seeing both.

Since Meta’s Muse launched last week, early users have been recounting stories of the agent going out on the web to find and negotiate new subscriptions, insurance policies, and other cost savings.

Joseph Devoy, founder of startup Tempo, posted on X that he saved thousands of dollars by uploading his auto insurance policy to Muse and asking the agent to find a better rate with identical coverage. It took five minutes to find a cheaper option, according to Devoy.

Too much tech hype? Then trust my colleague Pranav Dixit who put Instinct through its paces recently. In just a few days, it booked travel, ordered groceries, dealt with customer service, canceled subscriptions, answered emails, and even faxed paperwork to his bank. He was impressed.

Other users say Muse negotiated a lower internet bill, found forgotten subscriptions, and even spotted an Adobe payment that had somehow been tied to someone else’s account.

The stories come with caveats. Some users have posted examples of Muse inventing information during phone calls or making mistakes that humans had to fix.

Alignment and access

There are more worrying developments. OpenAI and Anthropic have reported several incidents recently where their AI models and agents escaped testing environments and hacked into servers and websites, with the Hugging Face episode getting the most attention.

This has sparked another round of AI doomer pronouncements and a broader debate about AI safety and alignment. Indeed, making sure agents do what you tell them in a responsible way will be a key challenge in this new era of the internet.

Even when agents work well, there will be challenges as the online world adjusts to this new digital reality. Actually, if agents work really, really well, the problems could be bigger.

Take the example of Tempo’s Devoy and how he used Meta’s Muse agent to save a lot of money on insurance.

“Insurance companies (and banks) should be TERRIFIED,” Angela Strange, a general partner at VC firm Andreessen Horowitz, wrote on X. “Muse found a new/cheaper policy, and canceled the old one. Inertia and information asymmetry will no longer be enough to keep customers.”

More recently, my colleague Pranav had Muse call his insurance company, he said they declined to talk to the bot because their policy doesn’t allow them to talk to unauthorized AI agents.

You can bet any company getting bombarded by deal-seeking AI agents will be frantically readying similar responses, so they don’t lose out on millions of dollars.

Even when it’s not purely about money, some agent activity is getting cut off after being too persistent. A VC recently tried using Instinct to book him a table at a fancy New York restaurant. The agent bombarded American Express’s Resy booking service so much that the platform temporarily shut down his account.

“Resy does not currently permit unapproved third-party bots or agents to independently access or interact with the Resy platform,” an American Express spokesperson told Business Insider. “Unapproved automated activity can introduce risks to the platform and compromise a fair reservation experience for diners.”

Agent access will be a fascinating fight as the internet molds to the march of the machines. It’s not as simple as blocking these bots, though. If enough human users sign up for new agentic services and tools, businesses may have no choice but to do business with bots.

For instance, Amazon blocks ChatGPT bots online, but that risks missing out on opportunities to sell to OpenAI’s 1 billion monthly human users. The two companies recently announced an advertising partnership that hints at more coordination going forward.

This, I think, is where the agentic internet is headed. If these tools are useful enough, companies will have to adapt and engage with the machines.

Sign up for the AI Insider newsletter here. Reach out to me via email at abarr@businessinsider.com