DeepSeek Harness 团队晒 Agent 架构数据:批量合并让单决策摊薄推理时间降至 1.6 秒
teortaxesTex · x · 2026-11-01
teortaxesTex commented on the DeepSeek Harness team's published team-agent design: a solid but expected design, not yet paper material, not better than Dot and similar efforts—but the repo itself is interesting and he expects interesting results to grow out of it.
The attached figures detail the Router layer's batching: as decision-trigger arrival rate rises from 1 to 40 events/minute, average batch size grows from 1.0 to 2.8 while time per request only rises from 2.5s to 3.9s, so inference time amortized per decision drops from 6.4s to 1.6s. When bursts create a backlog, the Router combines queued events for higher effective throughput, and at current team size there's no need to partition the Router layer prematurely at the cost of global context continuity.
所属事件:DeepSeek 开源 Super-agent 原型，GitHub 热度破 2k 星(10 条相关)→
「编程与Agent」频道最新
- GitHub Copilot 云端体验正式上线：一次会话跨 Web/移动/CLI/Slack 通用 — 0xkarasy · 2026-11-01
- shadcn：AI 系统最终都会造两种 UI——Slack 已发生，Linear 是下一个 — shadcn · 2026-11-01
- 把 Claude 当同事而非聊天机器人：分阶段委派报告任务的实战方法论 — ClaudeWorkbench · 2026-11-01
- 开发者用 Grok Bot 语音控制 Tesla 导航与空调联动日历 — doooyle · 2026-11-01
- Toloka 榜单：GPT-6.1 Terra 多次通过率从 72% 跌至 59% — MParakhin · 2026-11-01
- Gage:MCP 服务器为 Agent 间交接加上签名封缄与人类验证 — crossware · 2026-11-01