Berkeley's Evolving Agent Teams beats FlashInfer: supervisor-reorganized coding agents hit 1.2x speedup
rohanpaul_ai · x · 2026-11-10
A UC Berkeley paper (arXiv:2610.36848) introduces Evolving Agent Teams, a multi-agent framework that escapes local optima where single coding agents stall on expert-written GPU kernels — Claude Code plateaued at 0.51x FlashInfer's speed on the MLA decode kernel.
Key design:
- A central supervisor that diagnoses each local optimum and restructures the agent team;
- A persistent file system (learning document, evolution history, per-team artifacts) grounding restructuring in accumulated experience.
Under matched token budgets it consistently beats strong coding-agent baselines. With an extended budget, it generates a CUDA kernel from scratch achieving 1.2x mean speedup over the production FlashInfer MLA paged decode kernel, winning 36 of 47 FlashInfer-Bench workloads. Also strong on KDA prefill (3.01x over vLLM on AMD MI350X, 3.44x on B200), MoE expert-offload serving on Apple silicon (30.3 tokens/s decode on a 48GB M4 Pro), and RoboDojo robot manipulation. Code is open source.
Related event: Berkeley's Evolving Agent Teams Beats FlashInfer in GPU Kernel Optimization(2 posts)→
More from coding & agent
- Inference, agents, databases: what software engineers must master as MCP turns SaaS into 'a database in a trench coat' — sh_reya · 2026-11-10
- Dev builds an entire MMO world in one day with Fable 5.2, saying stylized art is the real test for new models — TAbrodi · 2026-11-10
- Independent judge verifies Kurzgesagt clone: SSIM 0.9938, all 8 gates pass — RichardsonDx · 2026-11-10
- Agent clones 30s Kurzgesagt video in pure code: 720 frames, 99.4% frame match — RichardsonDx · 2026-11-10
- 'Vibe coded' becomes an insult — Willison, DHH and Linux kernel converge on one rule: own the review — alvelda · 2026-11-10
- Can an LLM rebuild a closed-source game from just inputs and pixels? — TheZachMueller · 2026-11-10