Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

AgentFleet reports cross-device scheduling of up to 377 tasks/s

AI summary

A report on AgentFleet claims its in-memory scheduler handled 200 tasks at 302 tasks/s and 1,000 tasks at 377 tasks/s, while a two-device WebSocket test reached 96 tasks/s. The author also reports automatic reassignment without task loss during device disconnects, including a 1,000-task test across eight devices and 24 agents. These are self-reported results, and the report says only the Grok adapter is currently stable for direct connections while the other 12 platform adapters remain skeletal.

In-depth insight

The report provides a reproducible benchmark command and source code, giving users a way to evaluate AgentFleet's cross-device orchestration and failover claims for themselves.

上次写了 AgentFleet 的架构(把每台设备编入一支 AI 舰队,跨设备协同)。这次用实测数据说话——「跨设备调度」到底是口号还是能力?

三个可复现的数字

1. 调度吞吐 303~377 tasks/s 调度内核 benchmark(4 设备 × 2 agent,纯内存调度):200 任务 302 tasks/s,1000 任务 377 tasks/s。真实多设备链路(2 设备 × 2 agent,真实 ws):96 tasks/s。

2. 设备掉线 0 丢任务 故障转移 benchmark:注入设备掉线后任务自动重派,lost = 0。大规模场景 8 设备 × 24 agent、1000 任务、dev-0 在 200ms 时掉线,1000/1000 完成、lost=0。

3. 真实物理机 5/5 免登录跑通 3 台真实云服务器(Ubuntu + 2 台 Windows)跑 codex/claude,全程免登录(codex 直连 kilo 免费源、claude 走本地 modelgateway anthropic 转换),5/5 completed。

为什么值得看

  • 修了一个 O(n²) 调度退化,吞吐从 13 涨到 414 tasks/s(约 30 倍)。
  • 秒级断连感知 + running→queued 自动迁移——这是单机工具结构上做不了的分布式容错。
  • 数字都可复现:npm run benchmark --workspace backend。

诚实说明

当前只有 grok 适配器稳定直连,其余 12 平台是骨架(安装/发现/网关路由可用,headless 直连待逐平台实测),欢迎共建。

仓库:github.com/narkco-kang… (MIT)

#AgentFleet#Codex#Claude#Grok

Original · 1 min readRead the original
Group chat

No comments yet. Start the conversation.