AutoKernel tunes GPU kernels with agents

This is the most credible version of "agentic code optimization" because the feedback loop is brutally concrete: compile, benchmark, accept or revert. The real bet isn't that agents write perfect kernels on the first try; it's that enough cheap, disciplined iterations can outpace human tuning.

• –The `keep/revert` loop plus fixed benchmarks is exactly how you keep an optimization agent from hallucinating wins.
• –Supporting both Triton and CUDA C++ makes it useful for teams at different maturity levels, not just Triton-native shops.
• –The Amdahl's-law scheduler is smart product design: it stops wasting time on kernels that can no longer move end-to-end latency much.
• –The 5-stage correctness checks and roofline analysis are more convincing than raw TFLOPS screenshots.
• –AutoKernel fits a broader wave of autonomous kernel work, but its open-source repo makes it easier for developers to inspect and extend.