started · updated
UC Berkeley researchers develop evolving AI agent teams for GPU optimization
Researchers at UC Berkeley have introduced a multi-agent system called “Evolving Agent Teams” that optimizes GPU kernels by dynamically restructuring its own organization. According to a research paper, a supervisor agent monitors the progress of coding agents and identifies when performance plateaus. When progress stalls, the supervisor reshuffles roles, coordination methods, and information sharing protocols to overcome obstacles.
The system utilizes a persistent file system to maintain institutional memory, allowing agents to learn from previous iterations and failures. In testing, the evolving teams achieved a geometric mean speedup of 1.21x over the production FlashInfer kernel across 47 workloads, outperforming it in 36 cases.
When compared under matched token budgets for MLA paged decode tasks, the evolving teams reached 0.74x the performance of FlashInfer, while single agents like Claude Code reached approximately 0.51x. The research also demonstrated significant performance gains on different hardware, including a 3.01x speedup on AMD MI350X and a 3.44x speedup on NVIDIA B200 in specific kernel tests.
Entities
AMD · Claude Code · FlashInfer · Nvidia · UC Berkeley