building for the internet's trillionth user
I remember my first experience using the Internet: it was at my childhood library. I’d use the public computers to do research for homework. Google existed. Online commerce, social networks, and streaming media did not.
Then came Neopets, AIM, Myspace, Facebook, eBay, Stripe, and Shopify. The web became a place for individual expression, then a place for strangers to connect, then a place for the world economy to grow. Over the past 30 years, the web has become the rails for human expression and human extension. Today, we live our lives on the Internet as much as we do in-person. Humanity no longer exists without it; to imagine a society without the web evokes thoughts of oppression, decay, and isolation.
Until very recently, the Internet was exclusively designed for humans. We’ve spent the past 30 years designing its software for human consumption: designers create UIs to catch the human eye; marketers religiously A/B test website traffic to derive our collective human preferences; website frameworks optimize for browser compatibility and bandwidth throughput on our personal machines. In contrast, APIs, the primary abstraction designed for automations, cover a small fraction of all the reads, writes, uploads, and clicks that we humans do on websites. Humans have always been the users at the end of the line, so humans built for humans.
AI is changing that paradigm. Agents are now capable of taking on wide-ranging, complex tasks in almost any domain, including tasks that don’t have pre-defined APIs. They can do this today by using websites the same way that humans do: with a browser.
In fact, this is one of the most magical parts of AI: AI doesn’t need to wait for the world to rearchitect its systems for it. In contrast to every other technological shift that required investment to bridge the old world to new, AI comes backward compatible with the existing Internet rails we have today.
What will our lives look like when we’re augmented by autonomous inference, capable of handling anything we ask of it? We’re already seeing developers build this future on KERNEL’s platform. It’s one where:
All of this is possible, and all of this is already happening when you look at the emerging usage patterns of the Internet. The majority of traffic is shifting from human users to AI[1]. That momentum will only grow.
Building open web infrastructure—for AI and humans alike—is one of the most important opportunities of our time. How we design the systems that govern this will shape our future, and it’s something we talk a lot about at Kernel. It will define how future generations access information, create wealth, and govern. It will also determine who gets to participate. To achieve its potential, I believe this new age must be defined by the following:
The web evolved from a small set of design principles: anyone could publish, build a server, or propose new protocols. The earliest web was remarkably decentralized with no single company “owning” the web or determining who could participate. The result today, while not perfect, is a mostly open ecosystem that is much larger than any centrally designed intranet. These values will matter more as AI becomes a first class user.
The web’s current posture toward agents is adversarial by default. CAPTCHAs, IP blocking, and MFA are all built on the assumption that non-human traffic is malicious. That assumption made sense in the past, but it’s now lossy at best. AI today can circumvent detection; the blunt tools that are used slow down both people and their agents. The result is a bad experience for the users they were designed to protect.
Fraud and DDOS attacks are indeed growing challenges on the web, but we need clear definitions for what counts as a “legitimate agent,” not bigger, blunter tools.
At KERNEL, we define legitimacy as having received consent by the end user to act on their behalf. This is the simplest framework that passes the smell test; it lacks legalese or arbitrary gatekeepers. Put simply, AI agents that are delegates of end users are as legitimate as end users themselves; those without delegation or end website provisioning (from tools like robots.txt) are not.
We also need a way to distinguish legitimate agents from bad actors. The infrastructure that arbitrates this must clear a high bar: it needs to be low-friction enough that adoption requires little to no effort from end website operators; it must be decentralized such that no single company becomes a gatekeeper; and it must be sufficiently rigorous to be broadly trusted.
The next generation of legitimate AI agents will exist to act on our behalf. They’ll log into our bank accounts, submit expense reports, schedule doctor’s appointments, and purchase goods using permissions we’ve granted them. To end users, they should be explicit, traceable, and revocable. To end services, they should be recognizable and auditable.
The web has many protocols to represent authentication and authorization, but fragmentation makes it impossible for a single standard to solve this for every agent on every website. We need new infrastructure, layered on top of existing systems, that allows users to grant fine-grained, revocable permissions to agents[2]. That infrastructure should allow websites to verify agents and who delegated them. Finally, agent identity should be portable across the web, rather than created independently by every application.
This will benefit everyone involved. End users will be able to trust that agents act with the authority they’ve been given, and services will be able to distinguish legitimate traffic from abuse.
One of the Internet’s greatest strengths is also what makes it difficult to wrangle: no one designed it all at once. Websites, APIs, frameworks, and protocols were built by different people, at different times, for different purposes. The result is an ecosystem that is fragmented by design.
Fragmentation has historically been expensive. Integrating two systems meant writing translation layers, maintaining brittle code, or waiting for standards to emerge. AI materially changes that paradigm. Rather than requiring the world to converge on a single protocol, AI can understand, translate, and navigate disparate systems.
This doesn’t eliminate the need for standards, but the beauty of AI is that we don’t need to wait for the entire Internet to be rearchitected. The Internet has always evolved incrementally, and AI is capable of accelerating that evolution by bridging the gaps between the systems we already have.
Infrastructure, when done well, can play an important role in supporting this, from caching frequently traversed paths so agents become more reliable, to minimizing the inference required, to even democratizing agent-first knowledge about the web.
We are at the precipice of one of the largest technological shifts in our lifetimes, and the future is remarkably unconstrained. That freedom also carries a responsibility to lay foundations that carry us forward.
It’s how I think about what we’re doing at KERNEL. We provide infrastructure that makes the web accessible to AI, preserving the openness, interoperability, and universality that made the Internet successful in the first place. We give AI the tools to participate in the web that already exists, and we’ll help shape what’s yet to come.
We build our product through this lens too. In the past 16 months, our team has gone from just my co-founder, Raf, and me to a team of 20, powering developers who’ve launched tens of millions of browsers this year alone.
A few months ago, our team came together to codify our product north stars that drive us toward this future. They are aligned to what we believe the world needs and role that infrastructure should play:
Build the perfect web infrastructure for agents
Developers build on our platform today, but agents are increasingly the end users of our services. We aim to build features that are uniquely useful to agents and make them more powerful using us than any other alternative.
Define the technical frontier
We care about being the best-in-class technical platform for AI agents: the fastest and most reliable, at a cost that scales. We care about this because amazing infrastructure can unlock new use cases, expand the market, and are what agents at scale require. We also do it because it’s f*cking cool to push the boundaries of what is technically possible.
Empower legitimate agents
Web infrastructure is one of the only hot paths that sees every agent interaction on the web. We sit on the side of builders. Because we see the code and talk to our customers, we uniquely can understand an agent’s intent. Acting as a third party gives us the ability to distinguish legitimate agents and help other providers in the stack do so too.
Compound the capabilities of every agent on the web
We aim to make the web faster, easier, cheaper, and more reliable for every agent that uses it. There is lots that we can do at the infrastructure layer to make that possible. Over time, each agent on our platform makes the next one better, enabling an agent-first web experience built upon the rails of the original.
There’s a lot of history still to be written. When we succeed at building the next generation of the AI internet, however, I believe that the web will be more humane than ever: for humans and AI alike, used by both for the benefit of all.
==
[1] source: Agent traffic has surpassed human traffic as of June 2026
[2] KERNEL's Managed Auth product is our first step in this direction.
july 9, 2026
← back