---
title: "Evolving Agent Teams"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2610.36848v1"
permalink: "https://feed.astralor.com/items/itm_f7327ff09ebfbf1c3d46dd98"
published: "2026-10-27T08:46:21.000Z"
score: 57
channel: "tech"
tags: ["LLM","AI Agent","推理效率","Evolving Agent Teams","GPU","However","FlashInfer"]
---

# Evolving Agent Teams

## At a glance

The rapid progress of software engineering agents drives interest in applying them to open-ended problems such as GPU kernel optimization. However, we find that existing coding agents converge to local optima on real-wo…

## Key points

- The rapid progress of software engineering agents drives interest in applying them to open-ended problems such as GPU kernel optimization.
- However, we find that existing coding agents converge to local optima on real-world, expert-written kernels (e.g., FlashInfer).
- We introduce Evolving Agent Teams, a multi-agent framework that escapes such local optima.

## Full text

The rapid progress of software engineering agents drives interest in applying them to open-ended problems such as GPU kernel optimization. However, we find that existing coding agents converge to local optima on real-world, expert-written kernels (e.g., FlashInfer). We introduce Evolving Agent Teams, a multi-agent framework that escapes such local optima. It features (1) a central supervisor that diagnoses each local optimum and restructures a team of agents; and (2) a persistent file system that comprises a learning document, an evolution history, and per-team artifacts, grounding restructuring with accumulated experience. Under a matched token budget, Evolving Agent Teams consistently outperforms strong coding-agent baselines. With an extended token budget, Evolving Agent Teams generates a CUDA kernel that reaches 1.2x mean speedup over the production FlashInfer MLA paged decode kernel from scratch and wins on 36 of 47 production workloads from FlashInfer-Bench (0.86-0.99x on the remaining 11). It also demonstrates superior performance on tasks such as KDA prefill on AMD MI350X (3.01x over vLLM) and NVIDIA B200 (3.44x over FlashKDA), MoE expert-offload serving on Apple silicon (30.3 tokens/s decode for Qwen3.8-Flash-Next in MLX 4-bit on a 48 GB M4 Pro), and RoboDojo robot manipulation (62.1 on held-out seeds vs. 55.3 for the best leaderboard entry). Code is available at https://github.com/caoshiyi/EAT.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_f7327ff09ebfbf1c3d46dd98