How to use Cursor to build an experimental browser
Willison independently built Cursor’s agent-generated FastRender browser on macOS and found a working window with legible pages and visible rendering problems.
simonwillison.net
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
For [experimental browser repository], use planner agents to divide [browser goal] into worker tasks and a judge to assess each cycle. Then document a reproducible build for [platform] and inspect rendering of [test pages], recording glitches and build failures.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- srameshc
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- Willison obtained a working browser window and described tested pages as legible and mostly correct, with visible rendering glitches.
- Human involvement
- Willison independently built the browser on macOS and inspected Google and his own website, noting rendering glitches.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Planners and sub-planners created tasks for worker agents.
- Workers executed tasks with a judge assessing each cycle.
- Willison cloned FastRender and initialized its ecma-rs submodule.
Simon Willison examined Cursor's experiment in building a browser with a large fleet of coding agents. The reported coordination design used planners and sub-planners to create tasks, workers to execute them, and a judge to decide whether each cycle had completed the project. Cursor reported close to a week of agent work, producing over a million lines across 1,000 files.
Willison also tested the resulting FastRender repository himself. The initial announcement had attracted skepticism because GitHub Actions CI was failing and build instructions were missing. By his test, the README included instructions. On macOS, he cloned the repository, initialized the vendor/ecma-rs submodule and ran the browser through Cargo with the browser_ui feature.
That produced a working browser window. He captured Google and his own website and described the pages as legible and mostly correct despite obvious rendering glitches. The repository included WhatWG and CSS-WG specifications through Git submodules. Willison remained doubtful that such projects would compete with established browsers soon, while expressing surprise at the result.
Who it's for: developers investigating agent-built browser engines and independently checking experimental software.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Blog · srameshc
Frequently asked questions
Did Simon Willison run the generated browser?
Yes. He followed the updated README on macOS, initialized the ecma-rs submodule and ran the browser with Cargo, obtaining a working window.
How well did pages render?
He described Google and his own website as legible and mostly correct, while noting obvious rendering glitches.
More details
Remaining steps
- He launched the browser with Cargo and inspected rendered websites.
About this account
The reported Cursor fleet planned and implemented the browser across repeated worker and judge cycles.
Platform: Blog. Attribution: https://simonwillison.net/2026/Jan/19/scaling-long-running-autonomous-coding/. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: srameshc on Blog
Source published
Added to Agents In Use
Source checked
Tools and topics
Cursor · Git · Cargo · FastRender
Browser engine · Parallel agents · Rust
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.