Scaling long-running agents (FastRender)
Cursor reports hundreds of concurrent agents building a web browser from scratch, over 1M lines of code in about a week.
Planner/worker/judge hierarchy lets hundreds of agents work on one codebase for days. The write-up finds in-between structure works best and that prompts matter more than harness details; GPT-5.2 did best on sustained runs.
- Date
- Wednesday, 14 January 2026
- Lab
- Anysphere (Cursor)
- Kind
- paper
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| FastRender browser | >1M lines, ~1,000 files built in ~1 week per Cursor; Simon Willison's post (seen in search results, not opened) cites ~1.6M lines of Rust | company |
Research experiment, not a shipping product; Cursor's own account. Willison's 2026-01-23 FastRender post was not opened, so independent assessment of the result is not included.
Sources
This record was checked against its sources on 6 October 2026. How we check