Yeah we heavily leverage coding agents for optimizing our kernels. Since it's highly verifiable and takes time to measure we often leave multiple running and improving performance on different model architectures.

Definitely still helps to reference relevant academic work as well, or even just encouraging the agent to make bigger structural leaps, otherwise it will often get stuck working on low impact micro-optimizations.