Berkeley's Evolving Agent Teams beats FlashInfer by 1.2x with supervisor-run agent teams
burny_tech · x · 2026-11-05
UC Berkeley researchers introduce Evolving Agent Teams (EAT), a multi-agent framework that escapes local optima where coding agents get stuck optimizing expert-written GPU kernels.
- Core mechanism: a central supervisor diagnoses each local optimum and restructures the agent team on the fly (spawning debugging, profiling, or alternative-exploration agents), while a persistent file system (learning document, evolution history, per-team artifacts) grounds each restructuring in accumulated experience.
- Key results (extended token budget): a from-scratch CUDA kernel achieves 1.2x mean speedup over the production FlashInfer MLA paged-decode kernel, winning 36 of 47 FlashInfer-Bench workloads; 3.44x over FlashKDA on KDA prefill with NVIDIA B200 GPUs, plus wins on AMD MI350X (3.01x over vLLM), Apple silicon MoE serving (30.3 tokens/s decode), and RoboDojo manipulation (62.1 vs 55.3 leaderboard best).
- The paper's thesis: getting a kernel to run is not the same as making it fast — agents settle for polishing easier designs when a harder, hardware-specific path keeps failing.
Code is open-sourced at github.com/caoshiyi/EAT.
More from coding & agent
- Grok asked for the best coding harness recommends rival Claude Code — mhdfaran · 2026-11-05
- Agent benchmark scores hinge on container configs: 6-point swing from resource limits alone — krishnan · 2026-11-05
- Revolut cut AI coding inference costs ~90% by self-hosting GLM on 56 B200s instead of Claude — sebkrier · 2026-11-05
- VibeCoding a 3D fishing game in 4 days with Kimi K3.1 Pro and DeepSeek: full workflow breakdown — lxfater · 2026-11-05
- Agent eval design: testing partial updates vs full snapshots with a truck dock case — jylusdev · 2026-11-05
- Sentry founder says AI agent 'junior' handles hardest coding tasks well — zeeg · 2026-11-05