Scaling long-running autonomous coding

4 months ago

Wilson Lin at Cursor experimented with scaling autonomous coding agents, running hundreds concurrently on a single project.
The agents wrote over a million lines of code and trillions of tokens, using planners, sub-planners, and judge agents to coordinate tasks.
Cursor tested the system by building a web browser from scratch, resulting in 1 million lines of code across 1,000 files.
Initial skepticism arose due to failing GitHub Actions CI and missing build instructions, but these were later addressed.
The FastRender repo includes Git submodules for WhatWG and CSS-WG specifications, ensuring agents have reference materials.
A working browser window was achieved, showcasing impressive but glitchy rendering of web pages.
This achievement aligns with predictions of AI-assisted coding reaching complex projects like web browsers by 2029.
Another recent attempt at an AI-assisted browser is HiWave, a Rust-based browser engine.

Hasty Briefsbeta