A multi-agent coding system from UC Berkeley that restructures its own team mid-task produced GPU kernels faster than expert-written production code, achieving a geometric mean speedup of 1.21× over FlashInfer across tested workloads1,2.
The paper, titled Evolving Agent Teams and published on October 27, 2026, as arXiv:2610.36848, describes a supervisor agent that monitors a team of coding agents, diagnoses when progress has plateaued, and reorganizes roles and coordination on the fly.
The supervisor anchors its team around a persistent file system that stores learning documentation, a history of prior team configurations, and the artifacts agents produce.
ANALYSIS That shared memory lets lessons from earlier iterations carry forward rather than resetting between rounds. On a single task, the system explored over 100 configurations in search of better results.
The evolving teams came out ahead on 36 of 47 workloads tested. Applied to GPU kernel optimization, the approach produced code that beat expert-written production kernels on most of the workloads tested.
The results draw a sharp line between dynamic and fixed setups. On MLA paged decode tasks, fixed-role teams with a static structure managed 0.31× the performance of FlashInfer, while Claude Code, operating as a strong single agent, reached 0.51×.
ANALYSIS The gap between the static team at 0.31× and the evolving team's 1.21× geometric mean over FlashInfer suggests the reorganization mechanism itself accounts for a large share of the performance gain, rather than simply scaling agent count.
GPU kernels sit underneath nearly every AI model running today, and squeezing more speed out of them has traditionally been the job of a small group of very specialized human engineers.
ANALYSIS A system that automates the search over team structures and exceeds production-grade code on most workloads tested could shift how inference-stack performance work gets done.