XNews

Cursor Automates AI Research with Recursive Model Improvement and Agent Fleets

Cursor Automates AI Research with Recursive Model Improvement and Agent Fleets

Cursor's Lee Robinson presented their system for automating AI research to accelerate model iteration. The framework consists of an outer loop (user feedback to training) and an inner loop (improving training efficiency with harder RL environments and better reward models). A fleet of agents automates research tasks, allowing researchers to launch experiments from Slack and receive alerts when human intervention is needed. The key mechanism is recursive self-improvement: stronger models distill better auxiliary models, which in turn improve evaluation and data generation, accelerating the entire training loop. Cursor applied this in co-training Grok 4.5 with SpaceXAI, using massive real-world interaction data and difficult RL environments.

Source posts