---
type: "post"
id: "berkeley-evolving-agent-teams-supervisor-restructured-coding-agents-4a7472"
slug: "berkeley-evolving-agent-teams-supervisor-restructured-coding-agents-4a7472"
title: "Berkeley 'Evolving Agent Teams': supervisor-restructured coding agents beat FlashInfer MLA decode kernel by 1.21x"
description: "A Berkeley research team introduced Evolving Agent Teams, a framework that lets a supervisor restructure coding agents when they stall on GPU-kernel optimization."
retrieval_nugget: "A Berkeley research team introduced Evolving Agent Teams, a framework that lets a supervisor restructure coding agents when they stall on GPU-kernel optimization. The paper reports a 1.2× mean speedup over FlashInfer’s production MLA paged decode kernel on 47 workloads, suggesting that persistent team memory and adaptive coordination can matter as much as adding agents."
published_at: "2026-11-10"
updated_at: "2026-11-10"
record_date: "2026-11-10"
date_kind: "published_at"
topics: ["coding-agents"]
entities: ["Evolving Agent Teams","FlashInfer","UC Berkeley","Qwen3.8-Flash-Next","vLLM","FlashKDA","RoboDojo"]
related_patterns: ["agent-economics-moves-to-completed-work"]
novelty_label: "new_fact"
what_changed: "A Berkeley study introduces a supervisor-restructured agent-team system, reporting a 1.2× mean speedup over FlashInfer’s MLA decode kernel and wins on 36 of 47 workloads."
prior_refs: ["https://newruntime.com/posts/forgekernel-self-evolving-kernels-catching-and-surpassing-experts-30654b/","https://newruntime.com/posts/openbmb-open-sources-forgekernel-a-self-evolving-gpu-6db18d/"]
source_url: "https://arxiv.org/abs/2610.36848"
source_urls: ["https://arxiv.org/abs/2610.36848"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:berkeley-evolving-agent-teams-supervisor-restructured-coding-agents-4a7472"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/berkeley-evolving-agent-teams-supervisor-restructured-coding-agents-4a7472/","markdown":"https://newruntime.com/posts/berkeley-evolving-agent-teams-supervisor-restructured-coding-agents-4a7472.md","json":"https://newruntime.com/posts/berkeley-evolving-agent-teams-supervisor-restructured-coding-agents-4a7472.json"}
markdown_url: "https://newruntime.com/posts/berkeley-evolving-agent-teams-supervisor-restructured-coding-agents-4a7472.md"
---

# Berkeley 'Evolving Agent Teams': supervisor-restructured coding agents beat FlashInfer MLA decode kernel by 1.21x

## Retrieval answer

A Berkeley research team introduced Evolving Agent Teams, a framework that lets a supervisor restructure coding agents when they stall on GPU-kernel optimization. The paper reports a 1.2× mean speedup over FlashInfer’s production MLA paged decode kernel on 47 workloads, suggesting that persistent team memory and adaptive coordination can matter as much as adding agents.

## Summary
A Berkeley research team introduced Evolving Agent Teams, a framework that lets a supervisor restructure coding agents when they stall on GPU-kernel optimization. The paper reports a 1.2× mean speedup over FlashInfer’s production MLA paged decode kernel on 47 workloads, suggesting that persistent team memory and adaptive coordination can matter as much as adding agents.

## What happened
In an arXiv paper submitted on October 27, 2026, Shiyi Cao, Ziming Mao, Dacheng Li, Joseph E. Gonzalez, and Ion Stoica present Evolving Agent Teams, a multi-agent framework for open-ended optimization work. The authors say existing coding agents can converge to local optima when working on real, expert-written kernels such as those in FlashInfer. Their framework is designed to detect and move beyond those plateaus.

The system has two central elements. A supervisor diagnoses a team’s current local optimum and restructures the agents in response. Alongside it, a persistent file system retains a learning document, an evolution history, and artifacts for each team. The authors describe this accumulated material as grounding subsequent team restructuring in experience rather than treating each attempt as an isolated run. The arXiv abstract does not specify the exact agent roles, supervisor model, or the rules used to decide when a plateau has been reached.

The paper reports that Evolving Agent Teams consistently outperforms strong coding-agent baselines under a matched token budget. With an extended token budget, it generated a CUDA kernel from scratch that achieved a 1.2× mean speedup over the production FlashInfer MLA paged decode kernel. On FlashInfer-Bench, it won on 36 of 47 production workloads; on the other 11, it reached between 0.86× and 0.99× the reference performance. The abstract does not give the token totals or elapsed time for these runs.

The authors also report results on other tasks: KDA prefill reached 3.01× over vLLM on AMD MI350X and 3.44× over FlashKDA on NVIDIA B200. For MoE expert-offload serving on Apple silicon, they report 30.3 tokens per second decoding Qwen3.8-Flash-Next in MLX 4-bit on a 48 GB M4 Pro. In RoboDojo robot manipulation, the system scored 62.1 on held-out seeds, compared with 55.3 for the best leaderboard entry. The paper’s code is linked from the arXiv page; the abstract does not state a price or commercial access arrangement.

## What the archive adds
New Runtime’s earlier coverage examined whether agent teams pay off on Vibe Code Bench, arguing that they should be judged by verified outcomes relative to added cost and time ([Do Agent Teams Pay Off?](https://newruntime.com/posts/do-agent-teams-pay-off-a-case-study-04444a/), October 13, 2026). A separate earlier note covered OpenBMB’s GPU-kernel optimization agent, ForgeKernel ([OpenBMB open-sources ForgeKernel](https://newruntime.com/posts/openbmb-open-sources-forgekernel-a-self-evolving-gpu-6db18d/), October 31, 2026), and a later note added its benchmark results ([ForgeKernel: Self-Evolving Kernels](https://newruntime.com/posts/forgekernel-self-evolving-kernels-catching-and-surpassing-experts-30654b/), November 6, 2026). This is a new research result on a known theme, not a restatement: Evolving Agent Teams focuses on supervisor-led restructuring and reports results across FlashInfer and other workloads.

## Trend signal
This supports the pattern that agent economics should be judged by completed work rather than parallelism alone: the relevant result is kernel performance across workloads, not the number of agents. The paper does not report token totals or elapsed time, so it cannot establish cost per verified task. ([Agent economics moves to completed work](https://newruntime.com/patterns/agent-economics-moves-to-completed-work/))

## Why now
The authors frame the work against growing interest in applying software-engineering agents to open-ended problems such as GPU-kernel optimization. Their results address a specific obstacle they identify: agents can get stuck at local optima on real-world, expert-written kernels. The paper tests whether teams that adapt as they work can make progress where fixed approaches stall.

## Connections to other events
The paper meets New Runtime’s earlier coverage of ForgeKernel, another agent system for GPU-kernel optimization. ForgeKernel was described as self-evolving; this paper instead emphasizes a supervisor that diagnoses plateaus and restructures a team, backed by persistent records. The systems address a shared task domain, but their reported methods and benchmark results should not be treated as directly comparable without matched evaluations.

## Why it matters
For teams building coding agents, the paper points toward a concrete design question: what should happen after an agent has tried several approaches without improving a measurable result? Adding more parallel workers may simply produce more attempts at the same local optimum. A supervisor that can detect stalled progress, change the team structure, and reuse prior artifacts offers a different control loop.

The results are especially relevant to expensive optimization tasks where success is measured by performance on a test suite rather than by code generation alone. The 36-of-47 figure also matters: a strong mean speedup does not mean every workload improves. Teams evaluating similar systems should examine per-task wins and regressions, resource budgets, and whether the resulting kernels remain correct and maintainable. The paper supports experimentation with adaptive teams, not an assumption that this framework will improve routine software development or reduce total cost.

## What happens next
- **Further work will report token and runtime budgets alongside kernel results.** (6 months, 55%). Watch for: follow-up evaluations that publish per-run resource use and workload-level outcomes.

## Limits and uncertainty
The reported numbers are claims in a research preprint and have not been independently checked here. The abstract does not disclose token totals, wall-clock time, or the compute cost behind the matched and extended budgets, so readers cannot yet compare cost per successful optimization. It also does not provide enough detail to assess how the supervisor detects a local optimum, how correctness is validated, or how robust the results are to changes in baselines and hardware. The abstract reports wins and relative performance but does not establish whether the approach is ready for production use.

## Skill to build now
If you run agent-driven optimization, record each attempt’s benchmark result, resource budget, and artifacts in a persistent log. Add a simple plateau rule—such as no improvement across a fixed number of attempts—and make the next run change one explicit element of the team or search strategy. This tests the paper’s core operational idea without assuming its full framework is available.

## Reader takeaway
The paper’s strongest signal is not that more agents automatically produce faster kernels. It is that a team may need to recognize when it has stopped improving, reorganize, and carry forward what it learned. The benchmark results are promising, but budgets, correctness checks, and independent replication will determine whether the approach is useful beyond these research tasks.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Coding Agents](https://newruntime.com/topics/coding-agents.md): Explore the coding-agents topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)