By now, you know that AI can write code. Big names like Google and Microsoft have been mandating a certain percentage of their code to be written via AI. But how much can AI tolerate? And more importantly, will a mass-centric code written by AI work? Cursor’s CEO tested the theory to the limits. In a demonstration of how far AI has advanced in software development, Cursor CEO Michael Truell announced that a team of hundreds of AI agents, powered by OpenAI’s GPT-5.2 Codex model, autonomously constructed a full web browser in under seven days — producing more than 3 million lines of code without any human intervention.
The unusual experiment, which was shared by Truell on X (formerly Twitter), involved the AI agents running uninterrupted for a week to create a browser from the ground up. The project, which the company has referred to in related announcements as “FastRender,” features a completely custom rendering engine written in Rust. This includes core components such as HTML parsing, CSS cascading, layout engine, text shaping, painting, and even a custom JavaScript virtual machine.
Cursor CEO makes AI write 3 millions lines of code
Truell shared a screenshot showing the AI-built browser successfully loading the Google homepage. “It’s 3M+ lines of code across thousands of files. The rendering engine is from-scratch in Rust with HTML parsing, CSS cascade, layout, text shaping, paint, and a custom JS VM,” he stated.
However, he was quick to not let expectations run wild, admitting the browser only “kind of works” and remains “very far from WebKit/Chromium parity.” In follow-up responses, Truell clarified that the result was purely experimental, aimed at testing the limits of long-running autonomous coding agents rather than a production-ready product.
“There are definitely lots of issues!” he noted, highlighting that it is “very very far from production software.”
Internet gave mixed reactions
The experiment sparked widespread interest and debate online. While many in the tech community praised the results as a sign of rapidly evolving AI capabilities handling massive and complex projects, others expressed skepticism about its practical usability. For context, established browser engines like Chromium (the foundation for GoogleExplainer: How the Apple-Google pact changes the AI equation Chrome, Microsoft Edge, and others) contain over 35 million lines of code, built over decades by thousands of human developers.
Most reactions on social media included users requesting the exact prompts used to initiate the project, questions about the computational cost (in terms of Cursor tokens), and speculation on whether the AI drew inspiration from existing engines like Chromium. Some critics even labeled the output as ‘AI slop’.
Cursor, however, isn’t the first time to have done something like this. Anthropic’s Claude Sonnet 4.5 model has been claimed to code full applications in around 30 hours.