UC Berkeley proposes Evolving Agent Teams for GPU kernel optimization
UC Berkeley proposed Evolving Agent Teams, a framework that uses a central supervisor to reorganize coding agent teams as GPU kernel optimization encounters new bottlenecks. Worker agents handle kernel development, while the supervisor can add debugging, profiling, or knowledge-transfer roles and maintain shared storage for code, failures, hardware observations, and validated strategies. The team reported gains over strong coding-agent baselines in four CUDA kernel tasks, but each main experiment ran only twice, limiting statistical confidence.
The framework explores how dynamically restructuring roles and preserving shared knowledge may help AI agents pursue long-running GPU optimization tasks that otherwise settle into local optima.
GPU Kernel 优化,并不是简单地把一段可运行代码改得更快。面对真实生产环境中的 Kernel,coding agent 需要在硬件指令、内存布局、并行策略和整体架构之间反复探索,虽能在初期取得进展,但会逐渐转向更容易实现、性能上限较低的方案,最终陷入局部最优。
为解决这一问题,加州大学伯克利分校(UC Berkeley)团队提出了“持续进化 agent 团队”(EvolvingAgent Teams)。他们认为,与其增加 agent 数量,不如让中央监督器根据当前瓶颈,逐轮重组 agent 团队,调整角色分工、协作方式和知识管理策略。
论文链接:https://arxiv.org/abs/2610.36848
整个Evolving Agent Teams 框架分为两层:底层 agent 团队负责编写、编译、测试和分析 Kernel;上层监督器则观察优化进展,分析成功与失败的原因,并决定下一轮如何调整。
#Evolving Agent Teams#UC Berkeley#Claude Opus 4.6#Claude Code#Codex#MLA decode kernel#FlashInfer#ThunderKittens