Berkeley's Evolving Agent Teams beats FlashInfer: supervisor-restructured coding agents hit 1.2x speedup, win 36/47 workloads
rohanpaul_ai · x · 2026-11-10
A UC Berkeley paper introduces Evolving Agent Teams, a multi-agent framework targeting a known failure mode: coding agents make fast early progress on GPU kernel optimization, then settle on easy designs and plateau far below expert code — Claude Code stalled at 0.51x FlashInfer's speed on the MLA decode kernel.
The framework adds a central supervisor that diagnoses each plateau and restructures the team, inventing roles like profiler, debugger, and knowledge integrator. A persistent file system preserves lessons from every past team.
- Same token budget: 0.74x FlashInfer on MLA decode vs. 0.31x for a fixed 4-role team
- Extended budget: 1.2x mean speedup, winning 36 of 47 production workloads on FlashInfer-Bench
- Also strong on KDA prefill for AMD MI350X (3x over vLLM), MoE expert-offload on Apple silicon, and robot manipulation
Practical takeaway: for long, hard optimization runs, have a supervisor restructure agent roles when progress stalls instead of running one fixed setup. Code is open source.
Related event: Berkeley's Evolving Agent Teams Beats FlashInfer in GPU Kernel Optimization(2 posts)→
More from coding & agent
- Garry Tan's gbrain agent repo hits 32k GitHub stars — garrytan · 2026-11-10
- 8 concrete game-feel parameters to hand Claude: hitstop, flash, shake, with frame counts — algo_diver · 2026-11-10
- MARS trains RL agents to search raw corpora via shell commands, no retriever needed — _reachsumit · 2026-11-10
- Claude Plus Blender MCP Codes a 931-Part, 24,260-Rivet Jet Model — gaganghotra_ · 2026-11-10
- ENGRAM paper bakes agent experience into LLM weights, deletes memory only after verified tests — _reachsumit · 2026-11-10
- Tuned Claude skills win 95% of blind tests, beating one-shot prompt-written skills — every · 2026-11-10