In the GPU kernel addition budget experiment, 1.21x… on the same budget line, 0.75x

UC Berkeley researchers reported that dynamically restructuring AI coding teams improved GPU kernel optimization, achieving a 1.21x geometric-mean speedup over FlashInfer across 47 tasks when given extra token budget, though the gain fell to 0.75x under an equal-budget comparison. The work, detailed in the paper "Evolving Agent Teams" released October 27, 2026, tested MLA paged decode kernels used in large language model inference on NVIDIA H100 GPUs, with FlashInfer serving as the baseline for evaluating new kernels. The findings matter because they suggest agent-team organization, not just raw compute or token spend, can materially affect low-level GPU code performance for LLM inference. What to watch next is whether the equal-budget shortfall (0.75x) can be closed, determining if the approach offers real efficiency gains rather than only gains from additional token spend.