Photo: Pixabay / Pexels
A supervisor agent that reorganizes its team mid-task pushed GPU kernel performance past expert-written production code
A new UC Berkeley paper suggests AI coding agents work better with a manager who isn’t afraid to reshuffle the org chart.
The paper is titled “Evolving Agent Teams.” It describes a multi-agent system in which a supervisor watches a team of coding agents, spots when progress stalls, and restructures the team on the fly. Applied to GPU kernel optimization, the approach produced code that beat expert-written production kernels on most of the workloads tested.
That matters because GPU kernels sit underneath nearly every AI model running today. Squeezing more speed out of them has traditionally been the job of a small group of very specialized human engineers.
How the supervisor runs the team #
The supervisor agent monitors how the team is performing over a long coding task and diagnoses when the work has plateaued. When that happens, it changes who does what and how the agents coordinate with each other.
All of this is anchored by a persistent file system that the team shares. It stores learning documentation, a history of how the team has evolved, and the artifacts the agents produce. The result is a kind of institutional memory, so lessons from earlier iterations carry forward instead of evaporating between rounds.
According to the research, Evolving Agent Teams explored over 100 configurations on a single task in search of better results.
The numbers #
The headline result is a geometric mean speedup of 1.21× over the production FlashInfer kernel. The evolving teams came out ahead on 36 of 47 workloads.
The more telling comparison may be the one run under matched token budgets, where every approach got the same amount of compute to work with. On MLA paged decode tasks, the evolving teams reached 0.74× the performance of FlashInfer. Claude Code, operating as a strong single agent, reached 0.51×. Fixed-role teams with a static structure managed 0.31×.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
That last comparison is the most interesting finding. A static team of agents performed worse than a single agent working alone. Adding more agents without a way to reorganize them apparently created overhead rather than results.
The gains also extended beyond one chip vendor. The researchers reported speedups across hardware from NVIDIA, AMD, and Apple. On AMD’s MI350X, the evolving teams outperformed vLLM by 3.01× on KDA prefill tasks. On NVIDIA’s B200, they recorded performance of 3.44× over FlashKDA on additional workloads.
Who built it and where it stands #
The paper was published on October 27, 2026, as arXiv:2610.36848, and has been accepted as a poster at NeurIPS 2026. The project comes out of Berkeley’s Sky Computing Lab.
Shiyi Cao led the work, with contributions from Ziming Mao, Dacheng Li, Joseph E. Gonzalez, and Ion Stoica. The team has released the accompanying code on GitHub. Community discussion following the release has centered on the supervisory restructuring mechanism as pivotal for overcoming performance plateaus.
What this means for AI-driven coding #
The matched-budget results deserve the most attention from anyone weighing agent frameworks. Holding token budgets equal and still seeing the evolving team lead suggests the advantage comes from structure, not spending.
The static-team result is a useful caution. Teams building multi-agent products often assume that splitting a task among several specialized agents is automatically better than using one. On this evidence, a rigid division of labor can actually underperform a single agent.
The cross-hardware results widen the potential audience. GPU kernel tuning is typically done chip by chip, and expertise for one vendor does not always transfer to another. A system that produced gains on NVIDIA, AMD, and Apple hardware hints at a more general way to handle that porting work, though each result applies to the specific workloads and baselines the researchers tested.
There are limits worth noting. The strongest results compare against specific baselines on specific tasks, and the constrained-budget numbers still trail FlashInfer itself.
The persistent file system is another piece to track. Memory that survives across iterations is a design choice other agent frameworks could adopt, independent of the supervisor.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.