CUDA Agent and MaxKernel — agentic RL for accelerator optimization

CUDA Agent reports roughly 2x GPU-kernel speedups over compilers and prior models, while MaxKernel extends agentic compiler-feedback search to TPUs across 50 kernels and real workloads, claiming parity with experts. These results could enable software–hardware co-design, but benchmark transparency, independent reproduction, and generalization beyond kernel tuning remain unresolved.