Cursor's research team experimented with running hundreds of concurrent AI coding agents autonomously for weeks on large-scale projects. They evolved from flat self-coordination (which failed due to locking bottlenecks and risk-averse behavior) to a hierarchical planner-worker architecture. This system successfully built a web browser from scratch (1M+ lines of code), migrated Cursor's codebase from Solid to React (266K additions, 193K deletions over 3 weeks), and optimized video rendering 25x faster. Key learnings: GPT-5.2 excels at long-running tasks, different models suit different roles, simpler architectures outperform complex ones, and prompt engineering matters more than infrastructure. The approach demonstrates that multi-agent coordination can scale autonomous coding beyond single-agent limitations.
