GPU kernels are the small programs that run on graphics processors, and writing a fast one requires exploring many design choices. Today's best AI agents often get stuck in a mediocre solution and cannot find the best one.
Evolving Agent Teams is a system that uses several AI agents working together. A supervisor agent watches the team's progress and, when it gets stuck, reorganizes the agents with new tasks and knowledge from a shared file system. This lets the team change its approach as the problem evolves.
When applied to a widely used GPU kernel library called FlashInfer, the system achieved an average speedup of 1.21 times over the version written by human experts, beating it on most workloads. This shows AI can surpass human-written code in specialized domains, potentially reducing manual optimization effort.