YTGET

YouTube · Facebook · X · TikTok · Instagram · Threads

Google 開源 EGM 多代理審查流程,用金魚子代理挑錯並搭配平價模型省 token

YouTube 影片筆記

Google 開源 EGM 多代理審查流程,用金魚子代理挑錯並搭配平價模型省 token

原片標題:Google's NEW EGM Skill Set + Deepseek is ACTUALLY INSANE!

原片說明
Cline Desktop: https://cline.gg/desktop-aick ClinePass: https://cline.gg/clinepass-aick EGM on GitHub: https://github.com/google/elephant-goldfish-model PandaCode: https://github.com/aicodeking/pandacode In this video, I'll be taking a look at the Elephant-Goldfish Model, or EGM, an open-source plugin from Google that makes your coding agent design a feature properly before it writes any code, and that installs into Claude Code, Codex and Antigravity. I'll explain the idea behind it, where a full-context Elephant agent writes the design doc and fresh, zero-context Goldfish subagents review it, and then show how to install it and run the brainstorm, draft, review and resolve commands. After that, I'll show you why it pairs so well with Cline and the $9.99 per month ClinePass plan, how to mix models for the Elephant and the Goldfish, and how to run it inside PandaCode. Finally, I'll talk about what it means that Google built this for the harnesses where developers already are. -- Key Takeaways: 🐘 EGM is Google's open-source Elephant-Goldfish Model plugin: an Elephant agent with full context writes the design doc, and zero-context Goldfish subagents review it. 👀 Google built first-class install instructions for Claude Code, Codex, Claude Cowork and ChatGPT Work, not just for its own Antigravity agent. 🧠 Goldfish reviewers only see the design document and the files it points to, so they catch the shared assumptions you and your agent never wrote down. 💻 In Claude Code, every command uses the /egm: prefix: brainstorm sends out six Goldfish lenses in parallel, then draft, review and resolve take you to an approved design. ⚠️ The review skill bans flattery, and if your harness can't create fresh subagents, EGM stops the review instead of faking it in the same context. 🔍 Extra skills bootstrap README maps for an existing codebase and report drift between your design docs and the code. 💸 Five or six fresh reviewers burn a lot of tokens, so a $20 Claude or Codex plan will hit its limits quickly, and it's not meant for small bug fixes. ⚡ ClinePass is $9.99 a month with open models like Kimi, GLM 5.3 and DeepSeek V4.1 Flash, so you can run several review rounds and use different models for the Elephant and the Goldfish. 🔀 Cline supports SKILL.md skills and subagents, so you can run EGM in Cline or as read-only Cline workers inside PandaCode. 👍 Verdict: simple, thoughtful, free and Apache 2.0, and running it in Cline with ClinePass is a really good combo for $10 a month. -- Timestamps: 0:00 - Google's EGM plugin 0:43 - Google, but for Claude Code and Codex 1:33 - The idea Elephants and Goldfish 2:41 - Install and the brainstorm workflow 4:09 - Review, resolve and the no-flattery rule 5:42 - Bootstrap, drift report and the token cost 6:44 - ClinePass $9.99 a month and mixing models 8:27 - Running EGM in Cline and PandaCode 10:01 - Verdict the process matters more than the harness

Google 自己出了一個外掛,結果 README 裡第二條安裝指令是裝進 Claude Code,這畫面也太微妙了吧。我第一時間的反應是:你們不是有 Gemini 跟 Anti-Gravity 嗎?AICodeKing 這支影片就在拆這個叫 EGM 的開源專案,還有他怎麼把它塞進 Cline 搭配每月 9.99 美元的 ClinePass 跑。

Google 外掛卻先教你在 Claude Code 安裝
Google 外掛卻先教你在 Claude Code 安裝00:17

影片時間軸

  1. 01:39用大象與金魚的比喻說明兩種代理人的分工
  2. 02:43示範 Claude Code、Codex、Anti-Gravity 三種安裝指令
  3. 03:25跑 brainstorm,六個金魚評審並行挑毛病
  4. 03:55用 draft 分段寫設計文件,每段停下來等回饋
  5. 04:15review 五階段審查,從質疑架構到檢查術語
  6. 05:23規則明訂:生不出新子代理就直接中止審查
  7. 05:45bootstrap 與 drift report 兩個補充技能
  8. 06:48介紹 ClinePass 方案與混搭模型的玩法
五階段審查裡,金魚逐一挑出沒解釋的術語
五階段審查裡,金魚逐一挑出沒解釋的術語04:15

大象記住一切,金魚什麼都不記得

整個設計來自 Dave Renson 的文章,概念簡單到有點好笑:一個記住全部脈絡的大象代理負責寫設計檔案,再派一群零記憶的金魚子代理來讀。為什麼有效?你跟代理聊了一小時,雙方都累積了一堆沒說出口的共識,檔案讀起來順得很,因為缺的部分你們自己會補。換成一個什麼都不知道的讀者,他才會問「這個 webhook 掛掉怎麼辦」。這正是實作時會咬你的洞。

聊了一小時後,文件對雙方都讀得通
聊了一小時後,文件對雙方都讀得通02:12

不準誇獎,先講哪裡會出事

審查技能明文禁止評審說你的想法很棒,要求它們先講風險跟錯誤,還要給出最強的反對論證。如果你曾被代理稱讚一個明顯壞掉的計畫是「絕佳方案」,就知道這條有多重要。更硬的是:代理如果無法開新的子代理,EGM 直接停止審查,不在同一個脈絡裡假裝審。每個評審只拿到唯讀工具和獨立上下文,金魚才真的是金魚。

審查技能直接禁止評審講客套話
審查技能直接禁止評審講客套話05:01

代價是 token,所以配便宜方案

作者自己講得很白:小 bug 修復別用這套,設計加審查很花時間,小改動不一定回本。而且一輪審查就開五六個評審,每個都從頭讀文件跟參考碼,在 20 美元的方案上很快撞上限。於是他推 ClinePass,每月 9.99 美元,用 Kimi K3.1 Pro 當大象寫文件,GLM 5.3 當敵意評論者讀程式碼,輕量審查交給 DeepSeek V4.1 Flash。混模型還有額外好處:金魚連習慣跟盲點都跟大象不同。

金魚換成不同模型,連盲點都不一樣
金魚換成不同模型,連盲點都不一樣07:52

Google 把流程放到開發者所在的地方

作者認為這背後的故事更有意思:Google 等於承認代理的外殼沒那麼重要了。Claude Code、Codex、Anti-Gravity、Cline 都支援技能跟子代理,同一套流程寫一次到處跑,repo 也明說沒有任何外殼拿到分叉版本。價值從模型本身移到模型外面的流程,那把它放到開發者已經待著的地方就是合理選擇。對使用者來說,這代表流程不被綁在單一廠商,也不被綁在單一模型。

值得下載嗎

值得,如果你常跟代理做非小型的功能開發。它是免費的 Apache 2.0 專案,概念簡單、實作有原則,光是「禁止評審拍馬屁」跟「生不出新子代理就中止」這兩條就值得存下來參考。但記得先想好 token 預算,別在貴的方案上亂跑五階段審查。

覺得這篇筆記有幫助嗎?
YTGet 編輯室|看完影片整理的筆記,細節以原片為準。

連線中

影片封面

擷取畫面

輸入時間點,例如 3:20、3.20 或 200

影片皆已含音訊。檔案越大準備時間越久,完成前請保持頁面開啟。