NVIDIA open-sourced Polar, a reinforcement learning training framework that integrates existing code agents like Codex, Claude Code, and Qwen Code into GRPO training without modifying original code. It treats model API boundaries as training entry points, enabling black-box interception and trace reconstruction. In SWE-Bench tests, Codex pass@1 jumped from 3.8% to 26.4%. Training wall clock time shortened 5.39x with GPU utilization rising from 20.4% to 87.7%.