AI 代理人開一整隊,真的比只用一個好嗎? #GrokBot #ClaudeDuet #MetaMuse
Jim AI Notebook
0:00 / 0:00
AI 代理人開一整隊,真的比只用一個好嗎? #GrokBot #ClaudeDuet #MetaMuse
7 788 просмотров · 2 дн. назад
Jim AI Notebook
32,4 тыс. подписчиков
7 788 просмотров · 2 дн. назад
00:00 開場:GrokBot 的幕僚長與一整隊代理人
01:04 代理人是什麼、多代理人怎麼分工
03:13 Anthropic 研究系統:九成來自自己的考卷
04:02 一隊為什麼會贏:同時做、記憶乾淨、多一雙眼睛
05:08 帳單:贏有一大半是因為花得多
06:41 Walden Yan(Cognition):交接像傳話遊戲
07:48 Luke Alvoeiro(Factory):各畫半幅畫,合起來對不上
09:15 Yubin Kim(Google):分得開好八成,分不開差七成
10:50 16 個 Claude 寫編譯器
11:53 Noam Brown(OpenAI):同樣預算,一個不輸一隊
13:15 Simon Willison:最後的瓶頸是你的注意力
15:08 三個問題:分得開嗎?讀還是修改?值得嗎?
16:11 Lauren Tan(GrokBot):動手修改的只留一個
17:46 結論:先用好一個,再看要不要一隊
⚠️ 本片的原聲與畫面引用自各講者所屬公司或節目的官方頻道(YouTube 標準授權),作為評論與教育用途引用,每段約 30 秒以內,出處列在下方。Anthropic「分數高九成」是 Anthropic 自己內部研究題組的結果,不是公開基準;史丹佛的研究是尚未同儕審查的預印本;OpenAI 約一萬個代理人的數學成果仍有爭議。GrokBot 為 xAI 產品,訪談時 Lauren Tan 任職 Cursor(現屬 SpaceX)。本片不是投資或採購建議。製作流程有使用 AI 工具。
封面照片:Robonaut 2,NASA/Keegan Barber(公有領域)。
🎙 原聲出處
• Barry Zhang(Anthropic)|AI Engineer • How We Build Effective Agents: Barry Zhang...
• Sid Bidasaria(Claude Code 團隊)|Claude • How the Claude Code team uses Claude Code
• Walden Yan(Cognition)|Jason Liu 講座 • Why Devin (Cognition) does not use multi a...
• Luke Alvoeiro(Factory)|AI Engineer • The Multi-Agent Architecture That Actually...
• Yubin Kim(Google)|The AI Talks • [S5E7] Towards a science of scaling agent ...
• Noam Brown(OpenAI)|Dwarkesh Podcast • OpenAI researcher on agent swarms & recurs...
• Simon Willison|The Pragmatic Engineer • Simon Willison: Engineering practices that...
• Alex(Anthropic)|Databricks • How Anthropic Builds Multi-Agent Systems: ...
• Lauren Tan(GrokBot)|MTS • Are Agents About to Replace Software Engin...
📎 資料來源
• Anthropic〈How we built our multi-agent research system〉(2025 https://www.anthropic.com/engineering...
• Cognition(Walden Yan)〈Don't Build Multi-Agents〉(2025-06-12) https://cognition.com/blog/dont-build...
• Cognition〈Multi-Agents: What's Actually Working〉(2026-04-22) https://cognition.com/blog/multi-agen...
• Anthropic〈Building effective agents〉(2024-12-19) https://www.anthropic.com/engineering...
• Anthropic〈Effective context engineering for AI agents〉(2025- https://www.anthropic.com/engineering...
• Anthropic(Nicholas Carlini)〈Building a C compiler with a tea https://www.anthropic.com/engineering...
• Claude Code 官方文件:agent teams(實驗功能) https://code.claude.com/docs/en/agent...
• Kim 等〈Towards a Science of Scaling Agent Systems〉arXiv 2512. https://arxiv.org/abs/2512.08296
• Cemri 等〈Why Do Multi-Agent LLM Systems Fail?〉arXiv 2503.1365 https://arxiv.org/abs/2503.13657
• Tran & Kiela(Stanford)arXiv 2604.02460(預印本,2026-04):同樣思考預算下的 https://arxiv.org/abs/2604.02460
• OpenAI〈A practical guide to building agents〉 https://cdn.openai.com/business-guide...
• xAI〈Introducing Grok Bot〉(2026-08-11) https://x.ai/news/introducing-grok-bot
• Grok Bot 官方文件:Bots https://docs.x.ai/grok-bot/bots
• Quanta Magazine:OpenAI 約一萬個代理人的 Navier–Stokes 成果報導(2026-09-0 https://www.quantamagazine.org/
• AI Engineer〈How We Build Effective Agents: Barry Zhang, Anth • How We Build Effective Agents: Barry Zhang...
• Claude〈How the Claude Code team uses Claude Code〉(2026-09-02 • How the Claude Code team uses Claude Code
• Jason Liu〈Why Devin (Cognition) does not use multi agents〉Wa • Why Devin (Cognition) does not use multi a...
• AI Engineer〈The Multi-Agent Architecture That Actually Ships • The Multi-Agent Architecture That Actually...
• The AI Talks〈Towards a science of scaling agent systems | Yu • [S5E7] Towards a science of scaling agent ...
• Dwarkesh Patel〈OpenAI researcher on agent swarms & recursive • OpenAI researcher on agent swarms & recurs...
• The Pragmatic Engineer〈Simon Willison: Engineering practices • Simon Willison: Engineering practices that...
• Databricks Events〈How Anthropic Builds Multi-Agent Systems: • How Anthropic Builds Multi-Agent Systems: ...
• MTS〈Are Agents About to Replace Software Engineering? | Laur • Are Agents About to Replace Software Engin...
• Greg Isenberg〈Cursor AI Agents Work Like 10 Developers (Curs • Cursor AI Agents Work Like 10 Developers (...
• Lenny's Podcast〈How we built Grok Bot in a month | Roman Uga • How we built Grok Bot in a month | Roman U...
• Cursor〈Opening Keynote, Michael Truell | Compile 26〉(2026-06 • Opening Keynote, Michael Truell | Compile 26
• AI Engineer〈Don't Build Agents, Build Skills Instead – Barry • Don't Build Agents, Build Skills Instead –...
• Latent Space〈Anthropic vs Cognition on Multi-Agents: A Break • ⚡️Anthropic vs Cognition on Multi-Agents: ...
• SAIL Media〈Why Multi-Agent AI Fails (and How We Fix It) | Ne • Why Multi-Agent AI Fails (and How We Fix I...
• Human Feedback Foundation〈Paper Club with UC Berkeley: Why D • Paper Club with UC Berkeley: Why Do Multi-...
• Lenny's Podcast〈We replaced our sales team with 20 AI agents • We replaced our sales team with 20 AI agen...
• OpenAI〈Build Hour: Agents SDK〉(2026-05-28)(YouTube 標準授權,引用片段 • Build Hour: Agents SDK
---
Jim AI Notebook | Note the Future
每日深度解析 AI 最新進展
#AI代理人 #多代理人 #AIAgent #MultiAgent #ClaudeCode #GrokBot #Cognition #Anthropic #OpenAI #Cursor #AI工作流程 #JimAINotebook