Berkeley’s evolving agent teams beat expert GPU kernels by 1.21x
A UC Berkeley paper published October 27, 2026 (arXiv:2610.36848) describes a multi-agent system that writes faster GPU kernels than the expert-tuned production code underneath most AI models running today. The paper, titled Evolving Agent Teams, has been accepted as a poster at NeurIPS 2026.
How the supervisor works
A supervisor agent watches a team of coding agents over a long task. When progress plateaus, it restructures the team: changing roles, changing coordination patterns, and trying new configurations. The system explored over 100 configurations on a single task. A shared persistent file system stores learning documentation, team evolution history, and task artifacts, so lessons from earlier iterations carry forward rather than disappearing between rounds.
The numbers
The headline result is a geometric mean speedup of 1.21x over the production FlashInfer kernel, with the evolving teams winning on 36 of 47 workloads tested.
The more telling comparison used matched token budgets, where every approach got the same compute. On MLA paged decode tasks:
- Evolving teams: 0.74x FlashInfer performance
- Claude Code as a single agent: 0.51x
- Fixed-role static teams: 0.31x
That last number is the most useful caution for anyone building multi-agent systems. A rigid team of specialized agents performed worse than one agent working alone. Adding agents without a reorganization mechanism created overhead, not results.
Cross-hardware results extended the finding beyond a single chip vendor. On AMD’s MI350X, the evolving teams outperformed vLLM by 3.01x on KDA prefill tasks. On NVIDIA’s B200, they recorded 3.44x over FlashKDA on additional workloads.
What to take from this
The matched-budget results are the ones worth sitting with. The evolving team led even when token spend was held equal, which points to structure as the source of the gain rather than raw compute.
The persistent file system is also worth tracking as a standalone design choice. Memory that survives across iterations is something other agent frameworks could adopt independently of the supervisor mechanism.
There are limits worth noting. The strongest numbers compare against specific baselines on specific tasks, and the constrained-budget figures still trail FlashInfer itself. The code is available on GitHub. The work comes out of Berkeley’s Sky Computing Lab, led by Shiyi Cao with contributions from Ziming Mao, Dacheng Li, Joseph E. Gonzalez, and Ion Stoica.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
