TL;DR of the paper, generated automatically.

A coding agent that asks targeted questions beat a tool-heavy one on GPU kernels.

• Baseline: tool-rich agent stalled at 0.50x FlashAttention-3; this one reached 1.13x.
• Kernels: single CUDA files, Hopper and Blackwell, no CUTLASS or cuDNN.
• Speed: 1.16x FlashAttention-3 and 1.46x FlashAttention-4, geometric mean.