Cursor has published a detailed blog describing large-scale experiments where hundreds of AI coding agents were run autonomously for weeks at a time, pushing the limits of long-duration agentic software development.

In one experiment, Cursor organized agents into planners, workers, and judges—allowing them to collaborate in parallel—and tasked them with building a web browser entirely from scratch. The resulting project exceeded three million lines of code and was able to load simple websites correctly, all within under a week.

Additional experiments included building a Windows 7 emulator, an Excel-like spreadsheet application, and migrating Cursor’s own internal codebase. Each project spanned more than one million lines of code and required sustained autonomous execution rather than short interactive sessions.

The team noted that GPT-5.2 handled long autonomous runs more reliably than Claude Opus 4.5, which they observed was more prone to taking shortcuts during extended tasks. Cursor attributed the success to improved agent coordination, role specialization, and evaluation loops.

Why it matters: Coding agents are crossing an invisible threshold from task-level assistance to sustained, multi-week software construction. As agent coordination improves and autonomy windows lengthen, the economics and timelines of large-scale software development could shift fundamentally.