How to use Cursor to build a browser with parallel agents
Lin reported a peak of approximately 2,000 concurrent agents. FastRender displayed GitHub, Wikipedia and CNN pages, but JavaScript remained disabled and the project was a research experiment.
simonwillison.net
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Build an experimental [browser component] using planning and worker agents. Divide [scope] into minimally overlapping tasks, track unfinished functionality, and demonstrate rendering on [test pages]. Report limitations and coordination issues.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- lumpa
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- Lin reported a peak of approximately 2,000 concurrent agents. FastRender displayed GitHub, Wikipedia and CNN pages, but JavaScript remained disabled and the project was a research experiment.
- Human involvement
- Lin selected the experiment and demonstrated pages; the supplied excerpt does not describe a code review process.
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
- Started a browser rendering engine experiment with frontier models.
- Organized planning and worker agents into a task tree.
- Split browser work streams across separate harness machines.
Wilson Lin began FastRender as a personal experiment in whether frontier models could handle a difficult, well-specified software project. A browser rendering engine offered a large scope and a visible way to assess progress. After single agents made progress, the experiment became an official Cursor research project exploring cooperation among many agents.
The harness arranged agents in a tree: planning agents created tasks and workers carried them out. Lin split the browser into separate work streams, each running its own harness on a large machine. One demonstrated cluster handled CSS parsing and selectors. He reported that scope division minimized overlapping edits and that most commits avoided merge conflicts.
At the peak, approximately 2,000 agents ran concurrently during a week of continuous operation. The demonstrated browser loaded a GitHub repository page, Wikipedia and CNN, which were usable but somewhat slow. JavaScript remained disabled behind an agent-created feature flag while its engine was unfinished. Lin said the goal was to observe multi-agent behavior rather than deliver production software. Who it's for: researchers exploring large-scale coordination on complex software projects.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Blog · lumpa
Frequently asked questions
Was FastRender a production browser?
No. Lin said its purpose was to observe the behavior of multiple agents at scale, rather than build production software.
How did the agents avoid merge conflicts?
Lin said the harness divided scope to minimize overlap; most commits did not have merge conflicts.
More details
Remaining steps
- Demonstrated GitHub, Wikipedia and CNN pages in FastRender.
About this account
Planning agents assigned scoped tasks to workers; multiple harnesses ran browser work streams concurrently.
Platform: Blog. Attribution: https://simonwillison.net/2026/Jan/23/fastrender/. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: lumpa on Blog
Source published
Added to Agents In Use
Source checked
Tools and topics
Cursor · Multi-agent harness · FastRender
Parallel agents · Browser engine
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.