Cursor's research team built a multi-agent system capable of autonomous coding at scale, achieving ~1,000 commits per hour over a week-long run. The system evolved from single-agent failures through various coordination approaches to a final hierarchical design with planners, subplanners, and workers. Key learnings include
Table of contents
# Background# From single to multi-agent# Adding structure and roles# Continuous executor# The final system design# Throughput and tradeoffs# Infrastructure learnings# Specifying intent to agents# Optimizing prompts# System design learnings# ConclusionSort: