Understand this story
A team at UC Berkeley proposed the Evolving Agent Teams framework, in which an upper-level supervisor reorganizes the roles, division of labor, and collaboration methods of the underlying agent team round by round according to the current bottleneck, so as to break through local optima in GPU kernel optimization. Across four CUDA kernel tasks covering the NVIDIA B200 and H100, the framework outperformed strong coding agent baselines; under the same token budget, on the MLA decode task its performance reached 74% of the production-environment FlashInfer kernel, compared with about 51% for Claude Code, 31% for a fixed four-role team, and 28% for a single Codex agent. After investing more inference budget, the MLA kernel it generated achieved an average performance 1.21 times that of FlashInfer across 47 production workloads in FlashInfer-Bench, with 36 of those workloads exceeding the production-environment kernel. The team also noted that each of the four main experiments was run twice and most other scenarios were run only once, which is not yet sufficient to support a fully statistical comparison.
AI-generated from public reports
Latest development · Report
Berkeley team proposes Evolving Agent Teams, using dynamically reorganized agent teams to break through GPU…Developments
1 confirmed developments
- Source:36Kr · AI
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 30
36Kr · AI
Berkeley team proposes Evolving Agent Teams, using dynamically reorganized agent teams to break through GPU…
A team at the University of California, Berkeley proposes Evolving Agent Teams, in which an upper-level supervisor reorganizes the role division and collaboration methods of the underlying agent team round by round according to the current bottleneck, in order to break through local optima in GPU Kernel optimization. Across four CUDA Kernel tasks covering NVIDIA B200 and H100, the framework surpasses strong coding agent baselines; under the same Token budget, performance on the MLA decode task reaches 74% of the production-environment FlashInfer Kernel, compared with about 51% for Claude Code, 31% for a fixed four-role team, and 28% for a single Codex agent. After investing more inference budget, the MLA Kernel it generates achieves an average performance of 1.21 times that of FlashInfer across 47 production workloads in FlashInfer-Bench, with 36 of those workloads exceeding the production-environment Kernel. The team also notes that the four main experiments were each run twice and that other scenarios were mostly run once, which is not yet sufficient to support a fully statistical comparison.