Ask, Don't Equip: Question-Driven Coding Agents for Deep Kernel Optimization

Researchers introduce KernelSeek, a question-driven coding agent designed to optimize CUDA kernels without operator-specific knowledge or faster reference implementations. Using minimal harnesses and targeted questions at decision points, the agent outperformed tool-rich autonomous counterparts and produced dependency-free FlashAttention forward kernels for Hopper and Blackwell architectures. Open article →