Michael Truell, CEO of Cursor, dropped something wild on X this week.
His team built a browser. From scratch. In Rust. With a custom JavaScript VM.
The catch? They didn't write it. GPT-5.2 did.
Running uninterrupted. For one week. Across hundreds of agents.
3 million lines of code. Thousands of files. One experiment that changes how we think about AI coding.
The experiment
The project is called FastRender. It's open source now, so you can go look at what AI-generated browser code actually looks like.
Here's what the agents built:
- HTML parsing from the ground up
- CSS cascade and layout engine
- Text shaping and paint systems
- A custom JavaScript virtual machine
All in Rust. All from scratch. No forking Chromium. No borrowing from WebKit.
Does it work? Kind of.
Truell says simple websites render "quickly and largely correctly." It's nowhere near Chromium or WebKit parity. But the fact that it renders anything at all is the point.
Why GPT-5.2?
This wasn't random model selection. Cursor tested multiple options:
- Claude Opus 4.5 kept stopping early and asking for permission
- GPT-5.1-Codex had weaker planning capabilities
- GPT-5.2 could run autonomously for extended periods without losing focus
That last point matters. Long running tasks need a model that doesn't get distracted, doesn't give up, and doesn't keep asking "is this what you wanted?"
GPT-5.2 just kept building.
The architecture: planners, workers, reviewers
They didn't just point one agent at a blank repo and say "build a browser."
The system used a multi-agent hierarchy:
- Planners continuously explored the codebase and created tasks
- Workers focused on completing specific tasks
- Review Agents checked whether each iteration was good enough
This is the pattern. Not one super-agent. A coordinated system where different agents have different jobs.
This wasn't the only experiment
FastRender got the headlines, but Cursor ran other long-running projects too:
| Project | Lines of Code | |---------|---------------| | FastRender (browser) | 3,000,000+ | | Excel clone | 1,600,000 | | Windows 7 emulator | 1,200,000 | | JavaLSP | 550,000 |
That's over 6 million lines of AI-generated code across four experiments. In weeks, not years.
What this means
Chromium has 35 million lines of code. It took thousands of engineers decades to build.
FastRender has 3 million lines. It took a week.
Is FastRender going to replace Chrome? No. It's a proof of concept. A demonstration of what happens when you remove the human from the loop and let agents run.
But here's what I'm thinking about:
The gap between "kind of works" and "production ready" is getting smaller.
Every year, AI gets better at the boring parts: the glue code, the edge cases, the repetitive patterns. The stuff that makes projects take months instead of days.
We're not at "AI builds production software autonomously" yet. But we went from "AI can write functions" to "AI can build browsers" in about two years.
That trajectory matters.
The shift in how we use AI
Most people use AI like a fancy autocomplete. Write some code, ask for suggestions, accept or reject.
Cursor is testing a different model: set up the architecture, define the goals, let agents run for extended periods, review the output.
Less babysitting. More orchestration.
This is closer to managing a team than pair programming. You're not writing code together. You're defining what needs to happen and checking the results.
Try it yourself
The code is public: github.com/wilsonzlin/fastrender
Go look at it. Clone it. Run it. See what 3 million lines of AI-generated Rust looks like.
If you're building with AI agents, this is required reading. Not because the code is perfect. Because it shows what's possible when you push the boundaries of autonomous operation.
The real question
Everyone's asking "will AI replace developers?"
Wrong question.
The right question: "What can developers build when AI handles the volume?"
FastRender is 3 million lines in a week. Most of it probably needs refactoring. Some of it is probably wrong. But the scaffolding is there. The structure exists.
What would you build if you could generate a million lines of code as your starting point?
That's the question Cursor is answering.
FastRender is open source at github.com/wilsonzlin/fastrender. Michael Truell's original post is on X. Cursor's blog post on scaling agents is at cursor.com/blog/scaling-agents.
Discussion
Start the conversation by leaving a comment below.
No comments yet. Be the first to share your thoughts!