The RSI (recursive self-improvement) track continues to heat up, with AI systems shifting from objects of study to entities capable of self-development. At SOL-ExecBench, NVIDIA's global competition for GPU operator performance evaluation, the GPU operator optimization Agent framework built by domestic AI infrastructure vendor Infinigence AI took first place in four major tracks and ranked first among all participating teams, claiming 88 single-task first-place finishes. The framework establishes a complete iterative closed loop through execution boundary isolation, layered Skill and shared memory mechanisms, and a multi-abstraction-level design of operator implementation paths. Experiments also revealed that Agents competing for GPU resources can in turn undermine system efficiency, and that workflow evolution patterns differ across models. This practice explores two related paths for implementing RSI: first, efficient operators produced by Agents feed back into training and inference systems to improve computational efficiency; second, Agents participate in adjusting their own task execution mechanisms to optimize exploration and delivery capabilities. This confirms that AI infrastructure is not only the foundation for intelligent operation but can also become a core component participating in the continuous self-improvement of intelligent systems.