← back
Building Cursor Composer – Lee Robinson, Cursor
Original: Building Cursor Composer – Lee Robinson, Cursor
Takeaway
Cursor Composer trades raw frontier-model intelligence for 4x speed by RL-training on a production-mirroring environment with custom MoE kernels.
Summary
- Cursor Composer is a mixture-of-experts coding model trained via RL to be ~4x more token-efficient than peer frontier models at similar intelligence.
- Built with three servers (training/inference/environment) using Ray; load-balances rollouts that vary widely in tool-call count and latency.
- Custom low-precision kernels gave a 3.5x speedup on Nvidia Blackwell for MoE layers; written up in a Cursor blog post.
- Reuses Cursor's Cloud Agents VM fleet as the RL training environment so training mirrors production tool formats exactly.
- Parallel tool calling and semantic codebase search were the key features that pushed Composer over the daily-driver threshold internally.
coding-agentsreinforcement-learningcursor
Original description
Learn about the infrastructure, training, and evaluations used to build Cursor Composer, our first coding model. (https://cursor.com/blog/2-0) Speaker: https://x.com/leerob AIE is coming to London and SF! see https://ai.engineer for dates and sign up to be notified! **Timestamps** 00:00 Introduction to Cursor Composer 01:10 The "Fast vs. Smart" Trade-off 03:17 System Architecture & Tooling 04:33 Scaling Challenges: Consistency & Burstiness 05:50 Infrastructure Solutions & Custom Kernels 08:12 Co-designing Cloud Agents & Training Infra 09:39 The Power of Semantic Search 11:00 Results: Parallelism & Agent Behavior 12:13 The "Airplane Wi-Fi" Analogy 13:36 Key Reflections & Conclusion