Building Cursor Composer

Building agentic coding systems often fails on a familiar constraint: you can make them fast, or you can make them smart—but professionals need both to stay in flow. This talk walks through how Cursor built Composer, focusing on the infrastructure, training setup, and evaluations behind a low-latency coding agent model.

• Breaks down the “fast vs. smart” trade-off and why token-generation efficiency matters in real workflows
• Explains the rollout-based RL setup, including how tool calls (read/edit/search/lint/shell) are used and scored
• Covers scaling challenges like bursty compute, consistency between training and production, and load balancing for uneven rollouts
• Shows why matching the production environment—and integrating semantic search—shapes stronger agent behavior (e.g., better search/read before editing)