Multi-agent framework EAT hits 2.45x over SGLang on H200, SoTA across kernels and MoE serving
RulinShao · x · 2026-11-03
Evolving Agent Teams (EAT), a NeurIPS paper, introduces a multi-agent framework for long-running optimization tasks: a supervisor agent diagnoses why an agent team stalls and restructures it to escape local optima.
Reported results span multiple domains:
- Open Kernel Challenge: 2.449x over SGLang on decode layer (H200), ranked #4 of 1,180 as of Nov 1, 2026
- NVIDIA kernels: 3.44x over FlashKDA for KDA prefill on B200 (previous best: 2.96x)
- AMD kernels: 3.01x over vLLM for KDA prefill on MI350X
- MoE serving on Mac: Qwen3.8-Flash-Next on M5 Pro, 3.3x faster prefill (930 tok/s) and 1.22x faster decode (43.7 tok/s) than Strata at the same memory
- RoboDojo: 62.1 on held-out seeds
The authors position EAT as simple, effective and extensible, with SoTA performance across production kernels and heterogeneous hardware.
Related event: Berkeley's Evolving Agent Teams Set New Kernel Optimization Records(3 posts)→
More from coding & agent
- MoE leaderboard shows two spikes: ~300 teams hit 2.4x using Vjeux's public kernels — Vjeux · 2026-11-03
- User says Codex wiped systemd and containers instead of rebooting, costing $10,000+ — TheZachMueller · 2026-11-03
- Coder pitches an 'AI Operating Layer' to govern agents without losing control — CodeByPoonam · 2026-11-03
- DHH: programmers may be worse at AI coding than non-coders — '100%' — 2C_ornot2C · 2026-11-03
- Running a family trip with two agents: Teamily AI in action — thetripathi58 · 2026-11-03
- Dev says AI coding is collapsing to intent plus privileged SOR access — curious_vii · 2026-11-03