Improving Cursor’s agent for OpenAI Codex models

Cursor is enhancing its agent harness to integrate with frontier AI models, specifically focusing on OpenAI's latest Codex model, GPT-5.1-Codex-Max, to improve coding efficiency. This involves aligning tools and prompts with Codex's shell-oriented workflows, encouraging tool usage over shell commands for a better user experience, and implementing sandboxing for security. The team has refined reasoning summaries to improve user follow-along without causing distraction, and has introduced clear instructions for using linting tools to detect and correct errors in code. They have emphasized preserving reasoning traces to maintain continuity and performance, as their absence significantly impacts Codex's effectiveness. Additionally, instructions have been crafted to encourage autonomous action by the model, minimizing unnecessary user prompts, while message ordering is fine-tuned to prevent conflicts between system and user instructions. With the rapid pace of model developments, Cursor is committed to optimizing each new model's performance within their platform.