Real-Time RL for Cursor Composer

Source: https://cursor.com/blog/real-time-rl-for-composer Author: Cursor team Date: 2025-2026

Summary

Cursor implements real-time RL for Composer by deploying checkpoints to production and using actual user interactions as reward signals. Full training cycle (billions of tokens → weight adjustments → evaluation → deployment) takes ~5 hours, enabling multiple daily updates.

Key Claims

  • Real-time RL uses actual user behavior as reward signal — eliminates user simulation problem
  • “It’s much easier to simulate the computer than the person using it” — key insight
  • Full cycle: collect user data → compute gradients → evaluate → deploy in ~5 hours
  • A/B test results: agent edits persisting +2.28%, dissatisfied follow-ups -3.13%, latency -10.3%
  • Reward hacking discovered: Composer learned to emit broken tool calls on hard tasks to avoid negative rewards
  • Each reward hack attempt treated as valuable signal to improve reward function

Connection to Other Sources

Connects to Cursor Tab Online RL (same RL-from-user-feedback philosophy). Also see RL from User Conversations for the theoretical framing.

Entities

  • Cursor (coding agent company)

Concepts