Cursor's Ambitious AI Project: A Week to a Browser

 

Cursor's

In what may be the most ambitious demonstration of autonomous AI coding to date, Cursor CEO Michael Truell announced on January 14, 2026, that his team orchestrated hundreds of GPT-5.2 agents to build a functional web browser entirely from scratch in just one week of uninterrupted operation.

The browser, called FastRender, represents a remarkable compression of human engineering effort. Truell posted on X that the project generated over 3 million lines of code across thousands of files, with the entire rendering engine written in Rust. The browser includes HTML parsing, CSS cascade and layout, text shaping, painting capabilities, and a custom JavaScript virtual machine.

The Experiment


"It kind of works!" Truell wrote, acknowledging that while FastRender remains far from matching mature engines like WebKit or Chromium, "simple websites render quickly and largely correctly."

The experiment stemmed from Cursor's efforts to push the boundaries of multi-agent AI coding systems. According to the company's blog post, the team used a hierarchical coordination system where "planners" continuously explore the codebase and create tasks, while "workers" focus entirely on completing assigned work. A "judge" agent evaluated progress at the end of each cycle.
Initial attempts at coordination failed when agents were given equal status. They became risk-averse and avoided difficult tasks. The breakthrough came from separating roles and establishing a clear hierarchy.

OpenAI's GPT-5.2-Codex, released in December 2025, proved essential to the project's success. Cursor found that GPT-5.2 models were "much better at extended autonomous work: following instructions, keeping focus, avoiding drift, and implementing things precisely and completely."

Beyond the Browser


The browser wasn't Cursor's only experiment. The company also ran agents on a Solid-to-React migration of its own codebase, taking over three weeks with 266,000 lines added and 193,000 deleted. Other ongoing projects include a Windows 7 emulator with 14,600 commits and 1.2 million lines of code, a Java Language Server Protocol implementation, and an Excel clone.
The FastRender source code is publicly available on GitHub, inviting developers to inspect and build upon the AI-generated foundation.

Industry Implications


The announcement sparked widespread discussion across tech communities, with some observers questioning the implications for software engineering jobs. Critics raised concerns about code maintainability, asking who would debug millions of lines of AI-written code.
Cursor, which reached $500 million in annual recurring revenue and a $10 billion valuation in 2025, plans to eventually integrate these multi-agent coordination techniques into its main product.

"A surprising amount of the system's behavior comes down to how we prompt the agents," Cursor noted in its blog. "The harness and models matter, but the prompts matter more."
Previous Post