Cursor CEO Built a Browser using AI, but Does It Really Work?

‍

A new experiment by Cursor CEO demonstrates that AI can now autonomously coordinate massive coding projects, even building a browser.

Browser made using AI Agents

Michael Truell, the 25-year-old CEO of Cursor, and his team coordinated hundreds of GPT-5.2 agents to build a fully functional web browser from scratch.

AI agents worked “uninterrupted” for an entire week. The result? Over 3 million lines of code across thousands of files.

The secret was OpenAI’s GPT-5.2, a model specifically designed for extended autonomous work. GPT-5.2 was praised for its AI coding capabilities when it was released in December 2025. 

According to Truell, GPT-5.2 excels at precisely what this project demanded: staying on task, implementing features completely, and working autonomously for extended periods. 

The project was to build a fully new web browser, with its own rendering engine and core subsystems. 

The hardest part was getting hundreds of AI agents to work together without everything turning into a mess.

Initial attempts failed. When they gave all agents equal status and let them self-coordinate, the AI agents became paralyzed by indecision or became so risk-averse that they only made trivial changes.

Things started working when they separated roles:

• Planners identified what needed to be built and created task lists.
• Workers focused entirely on completing assigned tasks. 
• Judges evaluated progress and determined whether to continue for another iteration.

This hierarchical system solved the coordination problem.

In the end, they built a web browser from scratch, a project that typically takes experienced human developers months or even years to complete. The source code is publicly available on GitHub.

The codebase includes fundamental browser components such as HTML parsing, CSS cascade and layout, text shaping, painting mechanisms, and even a custom JavaScript virtual machine. 

Why is it getting appreciated?

A web browser might sound simple, but it’s one of the most complex pieces of software ever created. 

Modern browsers:

• render HTML and CSS
• execute JavaScript
• handle security protocols
• manage memory efficiently
• process network requests
• Display graphics

All while remaining fast and stable. 

The fact that AI agents could coordinate to build something this sophisticated is an achievement.

Although the browser remains far from matching Chromium or WebKit, Truell said it already renders simple websites quickly and largely correctly.

We need to understand that instead of stopping, throwing errors, and waiting for some human to fix it, it kept going. It debugged the issues on its own and moved forward.

Another interesting finding here was that GPT-5.2 outperformed Claude by keeping focus on multi-stage and complex tasks without human supervision.

But does it really work?

After making agents work 24x7, still Truell has to say: “It *kind of* works!”.

‍

Traditional browsers contain tens of millions of lines of code and took years of collective human effort to mature. Chromium has over 35 million lines of code. This AI-made browser is not even not 10% of them.

So, the browser is far from being production-ready software.

While the AI managed to crank out a bunch of code, there is skepticism about how effective that code actually is. 

See, rendering a webpage is not the hard part. The real complexity of a modern browser lives in everything around it, including extensions, password managers, security, accessibility, crash handling, and thousands of edge cases. 

None of that information is available here.

There is a problem with sustainability, too. Writing a million lines of code is one thing, but maintaining it is another. Fixing the remaining issues may take far longer than writing the initial code, in this case.

One developer on HackerNews even said trying to locate core components like the JavaScript engine or DOM implementation in the generated code is difficult. 

Also, real developers were needed for planning this whole thing. The agents didn’t spontaneously decide how browsers should be designed. Humans defined goals, roles, and workflows.

And don’t forget that browsers are one of the most heavily documented software categories. Even if no code was directly copied, the model has effectively been trained on decades of work. Is it just an advanced case of “copy and paste”?

Bottom Line

The real work begins when they need to remove the “kind of”. They might have stopped the “experiment” after one week because continuing further simply wasn’t worth the cost.

So, as of right now, such AI builds are still experiments. There is a lot for AI to learn to make complex software, and whether it will be able to do that or not is still a question mark.

Still, Cursor now plans to integrate these multi-agent coordination techniques into its main product soon.