Uploaded January 2026 | Updated November 2026, 1 day ago
Composer is a model built at Cursor for software engineering — trained with reinforcement learning and a mixture-of-experts architecture to deliver fast, agentic coding.
In this episode, Linda Haviv sits down with Sasha Rush, part of the team that built Composer, to talk about how the model was designed, why speed was a core requirement, and what it takes to build AI-native coding systems developers can actually trust and iterate with in practice.
Sasha walks through the vision behind Composer, the architectural choices behind its agent-based approach, and how reinforcement learning and mixture-of-experts enable specialization for real-world coding. We also discuss how Composer is trained and evaluated at scale using distributed infrastructure (including Ray), how developers are using it in practice, and what the future looks like for agent-based coding models and specialized AI systems.
This conversation focuses on how Composer was built and how it’s used, and complements Sasha’s Ray Summit keynote, which provides a broader overview of the work from Cursor’s research team.
Watch "Building Cursor Composer" - Ray Summit Keynote (Sasha Rush)
👉 youtu.be/md8D8eNj5JM?si=ABYNhlBsmfPvW19j
⏱️ Chapters
00:00 The Vision Behind Composer
00:31 Why Speed Matters for Coding Models
01:06 Architectural Choices: Reinforcement Learning and Mixture of Experts
01:55 Agent-Based Models and Tool Use in Coding
02:41 Training Composer at Scale: Infrastructure and Libraries
04:45 How to Use Composer in Practice
09:41 Future Trends and Advice for AI Model Development
11:33 Closing Thoughts on Agent-Based Coding
Composer is a model built at Cursor for software engineering — trained with reinforcement learning and a mixture-of-experts architecture to deliver fast, agentic coding.
In this episode, Linda Haviv sits down with Sasha Rush, part of the team that built Composer, to talk about how the model was designed, why speed was a core requirement, and what it takes to build AI-native coding systems developers can actually trust and iterate with in practice.
Sasha walks through the vision behind Composer, the architectural choices behind its agent-based approach, and how reinforcement learning and mixture-of-experts enable specialization for real-world coding. We also discuss how Composer is trained and evaluated at scale using distributed infrastructure (including Ray), how developers are using it in practice, and what the future looks like for agent-based coding models and specialized AI systems.
This conversation focuses on how Composer was built and how it’s used, and complements Sasha’s Ray Summit keynote, which provides a broader overview of the work from Cursor’s research team.
Watch "Building Cursor Composer" - Ray Summit Keynote (Sasha Rush)
👉 youtu.be/md8D8eNj5JM?si=ABYNhlBsmfPvW19j
⏱️ Chapters
00:00 The Vision Behind Composer
00:31 Why Speed Matters for Coding Models
01:06 Architectural Choices: Reinforcement Learning and Mixture of Experts
01:55 Agent-Based Models and Tool Use in Coding
02:41 Training Composer at Scale: Infrastructure and Libraries
04:45 How to Use Composer in Practice
09:41 Future Trends and Advice for AI Model Development
11:33 Closing Thoughts on Agent-Based Coding










