🌐 The Agentic Future of the Internet

Bots are back, baby!

The goal of the internet was to create a “collaborative space where you can communicate through sharing information” (Tim Berners-Lee). It's a simple goal with an endless depth of intricacy. Sharing information is a high bandwidth task. And the systems we’ve built to share it have become more and more complex. We've built interfaces that control power plants and tools that can spread messages to billions.

So much of our world now runs on the internet that people are the bottleneck! We’re the slowest part of the system, our hands dragging across the keyboard at 40 words per minute. The rate at which we can process, filter, improve, and react to new information is greatly inferior to the amount created every minute. Despite the ubiquity of modern software that greatly improves our efficiency, we still struggle to keep up.

That’s why this wave of AI is so exciting. It’s the panacea for the information overload we experience everyday. We now have a technology that can process information, understand it, and aid us in making decisions. People move one layer up the abstraction stack, from doing to supervising.

Much has been written about “agents” - essentially AI that can complete tasks for you. It pattern matches with every sci-fi movie you’ve ever seen. Your own AI in your pocket, going out into the world and submitting your expenses. The inspiring part of AI agents is that they’re unsupervised. You say what to do and they go do it.

But how will agents go complete tasks on your behalf? Well, until humanoid robots come along, agents are likely constrained to software. And since the majority of our tasks are completed on the internet, they’ll need to be plugged in.

There are over 1 billion websites are online right now and that number continues to grow rapidly. That means to be adequately effective, AI agents will need to use the web just like we do. As I’ve written before, AI browsing the internet is not an easy task, but it is a necessary one.

In the software world, “legacy code” is a scary word. Once something is working, and has been working for years, it becomes harder and harder to change it. All of the software we’ve built so far was designed for people, not agents.

This presents a challenge. If agents will be the primary users of software going forward, should we rewrite all software to integrate natively with agents? What if we still want humans to be able to use the software as well? Will we need to maintain two versions?

It seems easier and more practical to teach agents to use the web just as people do. Since 1/3 of the world still does not have access to the internet, it will take decades before agents become the main consumer of software.

Thankfully, we already have a common platform for interacting with the internet: the web browser. If you can teach an agent how to use a web browser, it unlocks the whole internet. That’s no small feat, and there are many simultaneous research initiatives focused on creating a reliable “web agent” (WebVoyager, SeeAct, Mind2Web).

It’s clear that the web browser will be a key piece of infrastructure in the agentic future of the internet. However, there are many challenges standing in the way: security, reliability, and speed.

That’s why we’re building Browserbase. Our vision is to be the internet gateway for all AI applications. By offering a programmable web browser designed for AI, we’ll enable developers to build powerful agents that 10x the productivity of humanity.

PS: We’re launching next week and would love your support. Make sure to follow along on Twitter and LinkedIn.