Evolving Agent Teams Outperforms Coding Baselines in CUDA Kernel Optimization

The Evolving Agent Teams framework uses a central supervisor and a persistent file system to prevent multi-agent teams from converging to local optima. In testing, it generated a CUDA kernel providing a 1.2x mean speedup over the FlashInfer MLA paged decode kernel.