Supervising and Reorganizing Coding Agent Teams Beats Expert GPU Kernels

Researchers at UC Berkeley found that introducing a supervisor that dynamically reorganizes coding agent teams when they stall on GPU kernel optimization significantly outperforms fixed-agent setups. Using a shared file system and inventing specialized roles, the supervisor helped agents reach 1.21 times FlashInfer speed under higher token budgets and beat expert code on multiple workloads. Open article →