EAT Evolves Agent Teams to Hit SOTA on Kernel Optimization, Ranks #4 in Open Kernel Challenge
berkeley_ai · x · 2026-11-03
A NeurIPS'26 paper introduces Evolving Agent Teams (EAT): a supervisor diagnoses why an agent team stalls and restructures it to escape local optima on long-running optimization tasks.
Key results:
- Open Kernel Challenge, decode layer on H200: 2.449× over SGLang, #4 of 1,180 teams
- NVIDIA B200 KDA prefill: 3.44× over FlashKDA (previous best 2.96×)
- AMD MI350X KDA prefill: 3.01× over vLLM
- MoE serving on Mac M5 Pro: 930 tok/s prefill (3.3× faster than Strata) and 1.22× faster decode
- RoboDojo: 62.1 on held-out seeds
The core idea is evolving the agent system's organization itself rather than prompts.
Related event: Berkeley's Evolving Agent Teams Achieve SOTA in CUDA Kernel Optimization(2 posts)→
More from coding & agent
- Model swap cut costs 59% and latency 75% — and he still blocked the release — Fun_Employment6042 · 2026-11-03
- Context drift broke a production LLM deployment by turn 10 — no jailbreak needed — GasProfessional5544 · 2026-11-03
- Obsidian dev open-sources Pele, a Mermaid-compatible renderer 100-500x faster than mermaid.js — Teknium · 2026-11-03
- Agent-built eval harness spawns subagents via codemode in one durable session — mitsuhiko · 2026-11-03
- SurfSense runs Claude Code and opencode as dual engines — lessons from a local NotebookLM — FurtiveMirth · 2026-11-03
- Impeccable crosses 100k GitHub stars, with weeks of releases teased ahead — jamdac · 2026-11-03