[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-claude-code-source-analysis-agentic-loop-zh":3,"tags-claude-code-source-analysis-agentic-loop-zh":35,"related-lang-claude-code-source-analysis-agentic-loop-zh":50,"related-posts-claude-code-source-analysis-agentic-loop-zh":54,"series-tools-7eb1b4a3-cdd0-4d33-8c95-5a0741de15cd":91},{"id":4,"title":5,"content":6,"summary":7,"source":8,"source_url":9,"author":10,"image_url":11,"keywords":12,"language":23,"translated_content":10,"views":24,"is_premium":25,"created_at":26,"updated_at":26,"cover_image":11,"published_at":27,"rewrite_status":28,"rewrite_error":10,"rewritten_from_id":29,"slug":30,"category":31,"related_article_id":32,"status":33,"google_indexed_at":34,"x_posted_at":10,"tweet_text":10,"title_rewritten_at":10,"title_original":10,"key_takeaways":10,"topic_cluster_id":10,"embedding":10,"is_canonical_seed":25},"7eb1b4a3-cdd0-4d33-8c95-5a0741de15cd","Claude Code 源碼拆解：五步循環與四層防護","\u003Cp>說真的，\u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fclaude-code\" target=\"_blank\" rel=\"noopener\">Claude Code\u003C\u002Fa> 這次被拆源碼，蠻有意思。文章裡直接點出 \u003Cstrong>五步 Agentic Loop\u003C\u002Fstrong>、\u003Cstrong>四層安全防護\u003C\u002Fstrong>、\u003Cstrong>三層 Agent 架構\u003C\u002Fstrong>。這不是一般聊天機器人的玩法。\u003C\u002Fp>\u003Cp>它更像一個會跑流程的系統。不是你問一句，它答一句。它會讀任務、規劃、呼叫工具、檢查結果，再決定下一步。講白了，就是把 LLM 放進可控的工作流裡。\u003C\u002Fp>\u003Cp>對台灣開發者來說，這種拆法很實用。因為我們平常不是在做 demo。真的上線時，卡住的常常是權限、上下文、測試、回滾。\u003Ca href=\"\u002Fnews\u002Fclaude-code-source-code-analysis-510k-lines-zh\">Clau\u003C\u002Fa>de Code 的設計，剛好碰到這些痛點。\u003C\u002Fp>\u003Ch2>先看整體：它不是聊天，是執行系統\u003C\u002Fh2>\u003Cp>很多人第一次看 \u003Ca href=\"\u002Fnews\u002Fopenclaw-1299-repos-eight-weeks-analysis-zh\">Cla\u003C\u002Fa>ude Code，會以為它只是命令列版 AI 助手。其實不是。它的核心是持續迭代的執行循環。這個循環把任務拆成幾段，讓模型每次只處理一小步。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775111420308-eo45.png\" alt=\"Claude Code 源碼拆解：五步循環與四層防護\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這種做法很像人類做專案。先看需求，再找資料，再改檔案，最後跑測試。差別在於，\u003Ca href=\"\u002Fnews\u002Fcloudflare-emdash-serverless-cms-wordpress-zh\">Cl\u003C\u002Fa>aude Code 把這些步驟都塞進系統裡，讓機器自己接著跑。\u003C\u002Fp>\u003Cp>這裡的重點，不是模型多會講。重點是它能不能把事情做完。對真實專案來說，能做完比會回答重要太多。\u003C\u002Fp>\u003Cul>\u003Cli>任務是循環式處理\u003C\u002Fli>\u003Cli>工具呼叫是流程的一部分\u003C\u002Fli>\u003Cli>適合多檔案、多步驟任務\u003C\u002Fli>\u003Cli>目標是完成工作，不只是回話\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>五步 Agentic Loop 怎麼跑\u003C\u002Fh2>\u003Cp>文章提到的五步流程，是理解 Claude Code 的關鍵。它大致會先整理任務，再做計畫，然後呼叫工具，接著讀回饋，最後決定要不要繼續。這種結構把黑箱拆開了。\u003C\u002Fp>\u003Cp>這樣做的好處很直接。哪一步出錯，一眼就能看出來。是計畫錯了，還是工具沒回應，還是驗證沒過。對工程團隊來說，這比純文字輸出好 debug 太多。\u003C\u002Fp>\u003Cp>我覺得這裡最猛的地方，是它把「思考」變成中間狀態。模型不是直接吐答案，而是一路修正。這很像一個會自己迭代的 junior engineer，只是速度快很多。\u003C\u002Fp>\u003Cp>另外，\u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Fclaude-3-5-sonnet\" target=\"_blank\" rel=\"noopener\">Claude\u003C\u002Fa> 系列的內部線索，也讓人看到它不是單一模型在撐場，而是整套系統一起配合。這種產品設計，比單純堆參數更像工程。\u003C\u002Fp>\u003Cblockquote>“The most important thing you can do is to make sure you have the right problem and the right solution.” — Dario Amodei, Anthropic co-founder and CEO\u003C\u002Fblockquote>\u003Cp>這句話放在 Claude Code 很貼。它不是在秀模型有多會寫，而是在處理一個更難的問題：怎麼把任務定義對，讓 AI 真能往下做。\u003C\u002Fp>\u003Ch2>上下文壓縮與記憶，才是長任務的核心\u003C\u002Fh2>\u003Cp>長任務最怕什麼？不是模型不夠聰明，是它忘東忘西。對話一長，歷史資料就會爆掉。模型會開始忘記約束，甚至前後打架。這是很多 AI 編程工具的老問題。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775111448458-ji56.png\" alt=\"Claude Code 源碼拆解：五步循環與四層防護\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>Claude Code 的解法，是做上下文壓縮和記憶管理。它不是死記所有內容，而是保留真正有用的部分。像任務目標、已完成動作、限制條件，這些會留下來。沒那麼重要的細節就壓掉。\u003C\u002Fp>\u003Cp>這個方向很務實。因為真實開發不是背誦聊天紀錄。真實開發要的是狀態連續性。系統知道現在做到哪，下一步該接哪，這才有用。\u003C\u002Fp>\u003Cp>文章也提到記憶系統不只是暫存。它會影響後續任務的偏好。這代表 Claude Code 不是只看當下對話，而是把過去經驗變成可重用上下文。這點很像一個會記得你專案習慣的同事。\u003C\u002Fp>\u003Cul>\u003Cli>壓縮無關歷史資料\u003C\u002Fli>\u003Cli>保留目標、限制與結果\u003C\u002Fli>\u003Cli>記憶會影響後續任務\u003C\u002Fli>\u003Cli>更適合長時間修改大型 repo\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>四層安全防護，才敢碰真實代碼庫\u003C\u002Fh2>\u003Cp>能進終端機，不代表能亂來。Claude Code 之所以能接進開發流程，安全設計一定要夠細。文章提到四層防護，這很合理。因為 Agent 一旦能讀寫檔案、執行命令，就不能只靠模型自覺。\u003C\u002Fp>\u003Cp>第一層是任務邊界。系統先決定它能處理什麼。第二層是工具權限。它能不能寫檔、跑指令、碰網路，都要管。第三層是行為檢查。危險操作要擋下來。第四層是審計與回饋。出事要能追。\u003C\u002Fp>\u003Cp>這種設計很像企業裡的權限控管。不是因為 AI 很壞，而是因為 AI 會犯錯。講白了，LLM 再強，也不能直接放飛。\u003C\u002Fp>\u003Cp>\u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fresearch\" target=\"_blank\" rel=\"noopener\">Anthropic Research\u003C\u002Fa> 一直強調安全與可控。Claude Code 的源碼拆解，把這件事講得很清楚。能執行的 AI，一定要有多道閘門。\u003C\u002Fp>\u003Cul>\u003Cli>任務邊界限制可做內容\u003C\u002Fli>\u003Cli>工具權限控制讀寫能力\u003C\u002Fli>\u003Cli>行為檢查攔高風險操作\u003C\u002Fli>\u003Cli>審計機制保留追蹤路徑\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>和其他編程助手比，差在哪裡\u003C\u002Fh2>\u003Cp>把 Claude Code 跟其他工具放一起看，差別很明顯。\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ffeatures\u002Fcopilot\" target=\"_blank\" rel=\"noopener\">GitHub Copilot\u003C\u002Fa> 很強，但它多半偏向補全和建議。\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fopenai\u002Fcodex\" target=\"_blank\" rel=\"noopener\">OpenAI Codex\u003C\u002Fa> 也很能生成，但常見形態還是圍繞輸出內容。\u003C\u002Fp>\u003Cp>Claude Code 更像任務執行器。它不是只幫你寫一段，而是追著任務往下做。這個差異很現實。前者適合局部補碼，後者適合跨檔案修改、跑測試、修錯誤。\u003C\u002Fp>\u003Cp>數字上也能看出差異。局部補全通常是 1 個檔案、1 段函式、1 次建議。Agent 化流程則常碰到 5 到 20 個檔案，還要經過多輪驗證。這兩種工作型態，根本不是同一件事。\u003C\u002Fp>\u003Cul>\u003Cli>Copilot 偏補全與提示\u003C\u002Fli>\u003Cli>Codex 偏生成與工作流\u003C\u002Fli>\u003Cli>Claude Code 偏跨檔案任務完成\u003C\u002Fli>\u003Cli>長任務裡，閉環比單次輸出重要\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>這其實是整個 AI 編程的方向\u003C\u002Fh2>\u003Cp>我覺得 Claude Code 的重點，不是某個神奇技巧，而是它把 AI 編程的方向講得很直白。未來的工具，不會只比誰寫得快。會比誰能理解任務、維持狀態、處理失敗，還能安全收尾。\u003C\u002Fp>\u003Cp>這也解釋了為什麼很多純聊天式產品，到了真實專案就卡住。因為真實專案不是一次問答。它是連續任務。你要改設定、補測試、修 CI、看 log、再改一次。沒有循環和記憶，根本撐不住。\u003C\u002Fp>\u003Cp>從產業角度看，這類 Agent 系統會越來越像開發流程的一部分。它們不只是 IDE 外掛，也不是單純對話框。它們會吃進 repo、吃進權限、吃進測試結果，然後自己往下跑。\u003C\u002Fp>\u003Cp>如果你問我，下一步該看什麼，我會看兩件事。第一，這類系統能不能在 10 分鐘內處理完整任務。第二，它在失敗時能不能自己回頭修。這比秀一段漂亮 code 重要多了。\u003C\u002Fp>\u003Ch2>結尾：別只看它會不會寫，先看它能不能做完\u003C\u002Fh2>\u003Cp>Claude Code 的源碼拆解，最有價值的地方，是把 AI 編程工具從「會講」拉到「會做」。五步循環、記憶壓縮、四層防護，這些東西拼起來，才像一個真的能進開發流程的系統。\u003C\u002Fp>\u003Cp>我的判斷很直接。接下來評估這類工具時，別先看 demo。先問三個問題：它能不能記住目標？能不能處理失敗？能不能在權限內安全收尾？如果答案都不差，那它就不只是助手了。\u003C\u002Fp>\u003Cp>你如果是工程師，現在就可以拿一個真實 repo 測。別拿玩具專案。拿一個有測試、有 lint、有 CI 的專案，看看它能不能真的把任務跑完。這才是最誠實的檢驗方式。\u003C\u002Fp>","從 Claude Code 源碼看五步 Agentic Loop、四層安全防線、三層 Agent 架構與記憶系統。這篇拆解它怎麼把聊天工具變成能跑長任務的開發流程。","zhuanlan.zhihu.com","https:\u002F\u002Fzhuanlan.zhihu.com\u002Fp\u002F2022442135182406883",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775111420308-eo45.png",[13,14,15,16,17,18,19,20,21,22],"Claude Code","Anthropic","Agentic Loop","LLM","AI編程助手","上下文壓縮","記憶系統","安全防護","終端機AI","源碼分析","zh",0,false,"2026-04-02T05:09:36.321443+00:00","2026-04-02T05:09:36.245+00:00","done","ef24b06d-2f27-46d0-896e-83384b9ea3d3","claude-code-source-analysis-agentic-loop-zh","tools","13d9b5c9-c800-4395-85ff-87ffb074b2a6","published","2026-04-09T09:00:52.081+00:00",[36,37,39,41,43,45,47,49],{"name":19,"slug":19},{"name":15,"slug":38},"agentic-loop",{"name":21,"slug":40},"終端機ai",{"name":17,"slug":42},"ai編程助手",{"name":13,"slug":44},"claude-code",{"name":14,"slug":46},"anthropic",{"name":16,"slug":48},"llm",{"name":20,"slug":20},{"id":32,"slug":51,"title":52,"language":53},"claude-code-source-analysis-agentic-loop-en","Claude Code 源码拆解：五步循环与四层防护","en",[55,61,67,73,79,85],{"id":56,"slug":57,"title":58,"cover_image":59,"image_url":59,"created_at":60,"category":31},"d058a76f-6548-4135-8970-f3a97f255446","why-gemini-api-pricing-is-cheaper-than-it-looks-zh","為什麼 Gemini API 定價其實比看起來更便宜","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778869845081-j4m7.png","2026-05-15T18:30:25.797639+00:00",{"id":62,"slug":63,"title":64,"cover_image":65,"image_url":65,"created_at":66,"category":31},"68e4be16-dc38-4524-a6ea-5ebe22a6c4fb","why-vidhub-huiyuan-hutong-bushi-quan-shebei-tongyong-zh","為什麼 VidHub 會員互通不是「買一次全設備通用」","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778789450987-advz.png","2026-05-14T20:10:24.048988+00:00",{"id":68,"slug":69,"title":70,"cover_image":71,"image_url":71,"created_at":72,"category":31},"7a1e174f-746b-4e82-a0e3-b2475ab39747","why-buns-zig-to-rust-experiment-is-right-zh","為什麼 Bun 的 Zig-to-Rust 實驗是對的","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778767879127-5dna.png","2026-05-14T14:10:26.886397+00:00",{"id":74,"slug":75,"title":76,"cover_image":77,"image_url":77,"created_at":78,"category":31},"e742fc73-5a65-4db3-ad17-88c99262ceb7","why-openai-api-pricing-is-product-strategy-zh","為什麼 OpenAI API 定價是產品策略，不是註腳","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778749859485-chvz.png","2026-05-14T09:10:26.003818+00:00",{"id":80,"slug":81,"title":82,"cover_image":83,"image_url":83,"created_at":84,"category":31},"c757c5d8-eda9-45dc-9020-4b002f4d6237","why-claude-code-prompt-design-beats-ide-copilots-zh","為什麼 Claude Code 的提示設計贏過 IDE Copilot","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778742645084-dao9.png","2026-05-14T07:10:29.371901+00:00",{"id":86,"slug":87,"title":88,"cover_image":89,"image_url":89,"created_at":90,"category":31},"4adef3ab-9f07-4970-91cf-77b8b581b348","why-databricks-model-serving-is-right-default-zh","為什麼 Databricks Model Serving 是生產推論的正確預設","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778692245329-a2wt.png","2026-05-13T17:10:30.659153+00:00",[92,97,102,107,112,117,122,127,132,137],{"id":93,"slug":94,"title":95,"created_at":96},"de769291-4574-4c46-a76d-772bd99e6ec9","googles-biggest-gemini-launches-in-2026-zh","Google 2026 最大 Gemini 盤點","2026-03-26T07:26:39.21072+00:00",{"id":98,"slug":99,"title":100,"created_at":101},"855cd52f-6fab-46cc-a7c1-42195e8a0de4","surepath-real-time-mcp-policy-controls-zh","SurePath 推出即時 MCP 政策控管","2026-03-26T07:57:40.77233+00:00",{"id":103,"slug":104,"title":105,"created_at":106},"9b19ab54-edef-4dbd-9ce4-a51e4bae4ebb","mcp-in-2026-the-ai-tool-layer-teams-use-zh","2026 年 MCP：團隊真的在用的 AI 工具層","2026-03-26T08:01:46.589694+00:00",{"id":108,"slug":109,"title":110,"created_at":111},"af9c46c3-7a28-410b-9f04-32b3de30a68c","prompting-in-2026-what-actually-works-zh","2026 提示工程，真正有用的是什麼","2026-03-26T08:08:12.453028+00:00",{"id":113,"slug":114,"title":115,"created_at":116},"05553086-6ed0-4758-81fd-6cab24b575e0","garry-tan-open-sources-claude-code-toolkit-zh","Garry Tan 開源 Claude Code 工具包","2026-03-26T08:26:20.068737+00:00",{"id":118,"slug":119,"title":120,"created_at":121},"042a73a2-18a2-433d-9e8f-9802b9559aac","github-ai-projects-to-watch-in-2026-zh","2026 必看 20 個 GitHub AI 專案","2026-03-26T08:28:09.619964+00:00",{"id":123,"slug":124,"title":125,"created_at":126},"a5f94120-ac0d-4483-9a8b-63590071ac6a","claude-code-vs-cursor-2026-zh","Claude Code 與 Cursor 深度對比：202…","2026-03-26T13:27:14.279193+00:00",{"id":128,"slug":129,"title":130,"created_at":131},"0975afa1-e0c7-4130-a20d-d890eaed995e","practical-github-guide-learning-ml-2026-zh","2026 機器學習入門 GitHub 實用指南","2026-03-27T01:16:49.712576+00:00",{"id":133,"slug":134,"title":135,"created_at":136},"bfdb467a-290f-4a80-b3a9-6f081afb6dff","aiml-2026-student-ai-ml-lab-repo-review-zh","AIML-2026：像課綱的學生實驗 Repo","2026-03-27T01:21:51.467798+00:00",{"id":138,"slug":139,"title":140,"created_at":141},"80cabc3e-09fc-4ff5-8f07-b8d68f5ae545","ai-trending-github-repos-and-research-feeds-zh","AI Trending：把 AI 資源收成一張表","2026-03-27T01:31:35.262183+00:00"]