[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-open-source-agent-stacks-seven-layers-2026-zh":3,"article-related-open-source-agent-stacks-seven-layers-2026-zh":32,"series-industry-2299155a-c8ca-40e1-9916-dd43f8f7a35f":79},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":24,"views":28,"created_at":29,"published_at":30,"topic_cluster_id":31},"2299155a-c8ca-40e1-9916-dd43f8f7a35f","open-source-agent-stacks-seven-layers-2026-zh","2026 開源 agent 堆疊拆成七層","\u003Cp data-speakable=\"summary\">O’Reilly 2026 指南把\u003Ca href=\"\u002Fnews\u002Fopen-source-ai-agent-frameworks-compared-langfuse-zh\">開源\u003C\u002Fa> \u003Ca href=\"\u002Ftag\u002Fagent\">agent\u003C\u002Fa> 工具拆成七層，幫團隊用生產需求挑工具。\u003C\u002Fp>\u003Cp>2026 年的開源 agent 世界，已經不是「選一個大框架」這麼簡單。Paolo Perrone 在 \u003Ca href=\"https:\u002F\u002Fwww.oreilly.com\u002Fradar\u002F\" target=\"_blank\" rel=\"noopener\">O’Reilly Radar\u003C\u002Fa> 2026 年 7 月 14 日的文章，直接把工具拆成七層。這個切法很務實，因為 demo 好看，和上線穩定，根本是兩回事。\u003C\u002Fp>\u003Cp>文章最狠的一點，是拿數字戳破幻覺。某個記憶框架在 \u003Ca href=\"https:\u002F\u002Flilys.ai\u002Fnotes\u002F1542\" target=\"_blank\" rel=\"noopener\">LoCoMo\u003C\u002Fa> 長對話記憶基準拿下好成績，卻可能比第二名重 340 倍。這種差距，足以改掉整個架構決策。\u003C\u002Fp>\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>訊號\u003C\u002Fth>\u003Cth>數值\u003C\u002Fth>\u003Cth>意義\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>文章日期\u003C\u002Ftd>\u003Ctd>2026-07-14\u003C\u002Ftd>\u003Ctd>放在 2026 年 agent 工具選型脈絡中\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>記憶基準重量差\u003C\u002Ftd>\u003Ctd>340 倍\u003C\u002Ftd>\u003Ctd>分數高，不代表生產成本低\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>主流 Python 專案星數\u003C\u002Ftd>\u003Ctd>50,000+\u003C\u002Ftd>\u003Ctd>顯示社群採用速度很快\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003Ch2>七層拆法，才是這篇真正的重點\u003C\u002Fh2>\u003Cp>Perrone 的核心觀點很直接。agent 堆疊已經分成七層：orchestration、memory、tool interface、browser 或 computer-use tools、coding agents、evals 與 observability、\u003Ca href=\"\u002Ftag\u002Finference\">inference\u003C\u002Fa>。這個拆法有用，因為每一層的瓶頸都不同。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784403178240-b3c3.png\" alt=\"2026 開源 agent 堆疊拆成七層\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>你如果把所有問題都丟給同一個框架，最後常常會卡在奇怪的地方。像是延遲太高、稽核軌跡不完整、\u003Ca href=\"\u002Fnews\u002Fkimi-k3-model-hype-into-harness-work-zh\">模型\u003C\u002Fa>綁死單一供應商，或是 Python 與 \u003Ca href=\"\u002Ftag\u002Ftypescript\">TypeScript\u003C\u002Fa> 團隊互相看不順眼。這些都不是 demo 會先告訴你的事。\u003C\u002Fp>\u003Cp>文章也把 \u003Ca href=\"\u002Fnews\u002Fopenclaw-v2026-7-1-control-ui-workspace-zh\">open\u003C\u002Fa> source 和 open core 劃清楚。若多租戶權限、複寫、SSO、audit log 只在雲端版，repo 再漂亮也不代表你能安心上 production。這點很現實，也很像工程師每天在踩的坑。\u003C\u002Fp>\u003Cul>\u003Cli>Latency budget 決定每輪能花多少時間和 Token。\u003C\u002Fli>\u003Cli>Audit trail 決定系統能不能被追查。\u003C\u002Fli>\u003Cli>Model portability 決定你會不會被單一模型綁住。\u003C\u002Fli>\u003Cli>Language stack 決定團隊能不能真的維護。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>Orchestration 層最容易先爆\u003C\u002Fh2>\u003Cp>在 orchestration 層，文章把 \u003Ca href=\"https:\u002F\u002Flangchain-ai.github.io\u002Flanggraph\u002F\" target=\"_blank\" rel=\"noopener\">LangGraph\u003C\u002Fa> 放在 Python 生產環境的預設位置。原因不是聲量，而是它處理 state 的方式很適合上線系統。durable execution、checkpointing、time-travel debugging，這些功能都很像為事故報告而生。\u003C\u002Fp>\u003Cp>如果你的系統要留痕，LangGraph 的設計就很對味。它的 \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Flangchain-ai\u002Flanggraph\u002Ftree\u002Fmain\u002Flibs\u002Fcheckpoint-postgres\" target=\"_blank\" rel=\"noopener\">PostgresSaver\u003C\u002Fa> 也讓狀態落地更自然，對台灣很多已經有 Postgres 的團隊來說，導入成本不算離譜。\u003C\u002Fp>\u003Cp>但它也不便宜。簡單的兩個 agent 流程，還是得先定 state schema、node、edge，再 compile。若你的需求只是順序呼叫三個工具，這套東西會顯得很重，像拿卡車送一個背包。\u003C\u002Fp>\u003Cblockquote>\u003Cp>“The best way to zero in on the constraint your system will hit first under load: latency budget, audit trail, model portability, or language stack.”\u003C\u002Fp>\u003Cfooter>Paolo Perrone, O’Reilly Radar, 2026-07-14\u003C\u002Ffooter>\u003C\u002Fblockquote>\u003Cp>這句話很準。它把 agent 選型從「哪個功能最多」拉回「哪個瓶頸先炸」。工程上，這才是比較像樣的問題。\u003C\u002Fp>\u003Ch2>輕量工具贏在小問題\u003C\u002Fh2>\u003Cp>這篇文章沒有把所有工具都往同一個天平上放。\u003Ca href=\"https:\u002F\u002Fwww.crewai.com\u002F\" target=\"_blank\" rel=\"noopener\">CrewAI\u003C\u002Fa> 的定位很清楚，就是低門檻。你定好 researcher、writer、reviewer 這種角色，它就能跑，對原型或內部試驗很快。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784403175044-0egg.png\" alt=\"2026 開源 agent 堆疊拆成七層\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>但快是有代價的。CrewAI 的 crash recovery 沒有做到 node 級別，錯誤處理也偏向 crew 層級。當你要做 postmortem，或要跟法遵、客服、SRE 對齊時，這種抽象就不夠細。\u003C\u002Fp>\u003Cp>\u003Ca href=\"https:\u002F\u002Fai.pydantic.dev\u002F\" target=\"_blank\" rel=\"noopener\">Pydantic AI\u003C\u002Fa> 則走另一條路。它把輸出綁到 typed Pydantic model，對單回合 agent 很友善。驗證和序列化都乾淨，接下游服務時少很多髒活。\u003C\u002Fp>\u003Cul>\u003Cli>\u003Ca href=\"https:\u002F\u002Flangchain-ai.github.io\u002Flanggraph\u002F\" target=\"_blank\" rel=\"noopener\">LangGraph\u003C\u002Fa>：適合需要 state 與重試的 Python 生產流程\u003C\u002Fli>\u003Cli>\u003Ca href=\"https:\u002F\u002Fwww.crewai.com\u002F\" target=\"_blank\" rel=\"noopener\">CrewAI\u003C\u002Fa>：適合快速做角色式原型\u003C\u002Fli>\u003Cli>\u003Ca href=\"https:\u002F\u002Fai.pydantic.dev\u002F\" target=\"_blank\" rel=\"noopener\">Pydantic AI\u003C\u002Fa>：適合 typed 單 agent 輸出\u003C\u002Fli>\u003Cli>\u003Ca href=\"https:\u002F\u002Fmastra.ai\u002F\" target=\"_blank\" rel=\"noopener\">Mastra\u003C\u002Fa>：適合 TypeScript 團隊和 Next.js 專案\u003C\u002Fli>\u003C\u002Ful>\u003Cp>TypeScript 團隊則常會看 \u003Ca href=\"https:\u002F\u002Fmastra.ai\u002F\" target=\"_blank\" rel=\"noopener\">Mastra\u003C\u002Fa>。它把 agents、workflows、\u003Ca href=\"\u002Ftag\u002Frag\">RAG\u003C\u002Fa>、evals 包在一起，對已經在 \u003Ca href=\"https:\u002F\u002Fnextjs.org\u002F\" target=\"_blank\" rel=\"noopener\">Next.js\u003C\u002Fa> 裡做產品的人很順。你不用先拉一個 Python sidecar，這點很實際。\u003C\u002Fp>\u003Cp>我覺得這裡最有價值的，是文章沒有硬推單一解法。團隊語言棧已經是 TypeScript，就別為了追風去硬上 Python。反過來說，如果你要的是可追查、可回放、可重試，Python 生態還是比較成熟。\u003C\u002Fp>\u003Ch2>分數好看，不代表成本漂亮\u003C\u002Fh2>\u003Cp>這篇最值得記住的，是它對 \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa> 的態度。分數高，不代表總成本低。記憶框架在 \u003Ca href=\"https:\u002F\u002Flilys.ai\u002Fnotes\u002F1542\" target=\"_blank\" rel=\"noopener\">LoCoMo\u003C\u002Fa> 上拿好成績，卻可能重 340 倍，這種結果會直接影響伺服器成本和延遲。\u003C\u002Fp>\u003Cp>同樣的邏輯也適用在其他層。browser tools 可能在簡單頁面很好用，但遇到 canvas-heavy 網站就開始出事。evals 如果只放在 Notion 裡，最後很容易變成大家都說自己有測，但誰也找不到測試紀錄。\u003C\u002Fp>\u003Cp>inference 層更不用說。你一旦把所有流程綁在單一模型供應商上，後面換模型的代價會很高。這也是為什麼文章一直強調要先看失敗模式，而不是先看排行榜。\u003C\u002Fp>\u003Cul>\u003Cli>記憶層：看的是延遲和記憶成本。\u003C\u002Fli>\u003Cli>瀏覽器層：看的是網站相容性。\u003C\u002Fli>\u003Cli>評測層：看的是可重現性。\u003C\u002Fli>\u003Cli>推論層：看的是模型切換彈性。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>這波工具分化，跟 2024 不一樣了\u003C\u002Fh2>\u003Cp>幾年前，很多團隊還在找「一個框架解決全部」的答案。現在不太行了。agent 工具已經分工，大家各做各的強項，這反而比較像正常軟體工程。\u003C\u002Fp>\u003Cp>對台灣團隊來說，這代表選型要更務實。金融、電商、客服、SaaS 的需求差很多。你要的是審計、重試、權限，還是快速出原型，答案會完全不同。\u003C\u002Fp>\u003Cp>所以這篇文章的價值，不在於它幫你選出唯一冠軍。它做的是把問題切小，讓你先問自己：哪一層最先壞？哪一層最貴？哪一層最難追？這三題答不出來，選工具多半也會選歪。\u003C\u002Fp>\u003Ch2>先看瓶頸，再決定堆疊\u003C\u002Fh2>\u003Cp>如果你現在要做 agent，我會直接建議先畫出失敗路徑。你的系統是怕延遲，怕資料外洩，怕模型鎖死，還是怕維護成本爆掉。這些答案會比社群聲量更有用。\u003C\u002Fp>\u003Cp>我的判斷很簡單。2026 年的開源 agent 市場，不會獎勵最花俏的 demo。它會獎勵能在 production 撐住的堆疊。先選層，再選工具，會少很多冤枉路。\u003C\u002Fp>","O’Reilly 2026 指南把 agent 工具拆成七層，並點名各層適合上線的開源方案，重點在生產環境的取捨。","www.oreilly.com","https:\u002F\u002Fwww.oreilly.com\u002Fradar\u002Fthe-open-source-agent-toolkit-in-2026\u002F",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784403178240-b3c3.png","industry","zh","4abfb46f-b476-4837-abe0-823deb9bef92",[17,18,19,20,21,22,23],"open source agent","LangGraph","CrewAI","Pydantic AI","Mastra","agent stack","2026",[25,26,27],"agent 工具在 2026 已經分成七層，不能再用單一框架思維選型。","benchmark 分數會騙人，340 倍記憶成本差距就是最直接的例子。","生產環境優先看 latency、audit trail、model portability 和 language stack。",0,"2026-07-18T19:32:32.446009+00:00","2026-07-18T19:32:32.434+00:00","778f6ce6-5165-4de6-a199-2d05555349fb",{"tags":33,"relatedLang":38,"relatedPosts":42},[34,36],{"name":18,"slug":35},"langgraph",{"name":19,"slug":37},"crewai",{"id":15,"slug":39,"title":40,"language":41},"open-source-agent-stacks-seven-layers-2026-en","Open source agent stacks split into seven layers in 2026","en",[43,49,55,61,67,73],{"id":44,"slug":45,"title":46,"cover_image":47,"image_url":47,"created_at":48,"category":13},"0c3a0e19-e3c7-48ae-8f7b-198ca0911957","llm-routing-benchmark-38-tasks-15-models-zh","5 款 LLM 的實戰路由結論","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784422962336-b8ht.png","2026-07-19T01:02:19.036094+00:00",{"id":50,"slug":51,"title":52,"cover_image":53,"image_url":53,"created_at":54,"category":13},"fb7e9fae-1b07-4f17-9445-62c0ae5ae401","openai-staff-fund-rival-super-pac-zh","OpenAI 員工捐 21.5 萬美元挺反 AI PAC","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784419368821-l8ql.png","2026-07-19T00:02:21.179629+00:00",{"id":56,"slug":57,"title":58,"cover_image":59,"image_url":59,"created_at":60,"category":13},"06a3244b-43b3-4735-86d1-d9d15de54c46","kimi-k3-model-hype-into-harness-work-zh","Kimi K3 把模型熱度變成編排工作","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784397788053-g4r8.png","2026-07-18T18:02:43.621005+00:00",{"id":62,"slug":63,"title":64,"cover_image":65,"image_url":65,"created_at":66,"category":13},"3afc4bba-0c4e-4fa9-9d30-8df2d6fc4792","laoyouji-di7ji-di12ji-zhihu-gao-zhi-shi-ban-yun-taici-zh","《老友记》第7季第12集：这篇知乎稿只是在搬运台词，不是内容","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784376191028-0830.png","2026-07-18T12:02:38.560235+00:00",{"id":68,"slug":69,"title":70,"cover_image":71,"image_url":71,"created_at":72,"category":13},"0ce05711-0dd4-42b8-b1a3-416fc11ba2f3","openai-should-pay-more-for-bio-jailbreaks-not-less-zh","OpenAI 應該為生物越獄支付更多，而不是更少","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784356377941-7f9e.png","2026-07-18T06:32:31.134944+00:00",{"id":74,"slug":75,"title":76,"cover_image":77,"image_url":77,"created_at":78,"category":13},"8537073f-636b-4239-a9da-21a810b0b183","cloudflare-q2-2026-earnings-call-aug-6-zh","Cloudflare 8 月 6 日公布 Q2 財報","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784316763172-r53n.png","2026-07-17T19:32:19.88221+00:00",[80,85,90,95,100,105,110,115,120,125],{"id":81,"slug":82,"title":83,"created_at":84},"ee073da7-28b3-4752-a319-5a501459fb87","ai-in-2026-what-actually-matters-now-zh","2026 AI 真正重要的事","2026-03-26T07:09:12.008134+00:00",{"id":86,"slug":87,"title":88,"created_at":89},"83bd1795-8548-44c9-9a7e-de50a0923f71","trump-ai-framework-power-speech-state-preemption-zh","川普 AI 框架瞄準電力、言論與州權","2026-03-26T07:12:18.695466+00:00",{"id":91,"slug":92,"title":93,"created_at":94},"ea6be18b-c903-4e54-97b7-5f7447a612e0","nvidia-gtc-2026-big-ai-announcements-zh","NVIDIA GTC 2026 重點拆解","2026-03-26T07:14:26.62638+00:00",{"id":96,"slug":97,"title":98,"created_at":99},"4bcec76f-4c36-4daa-909f-54cd702f7c93","claude-users-spreading-out-and-getting-better-zh","Claude 用戶更分散，也更會用","2026-03-26T07:22:52.325888+00:00",{"id":101,"slug":102,"title":103,"created_at":104},"bd903b15-2473-4178-9789-b7557816e535","openclaw-raises-hard-question-for-ai-models-zh","OpenClaw 逼問 AI 模型價值","2026-03-26T07:24:54.707486+00:00",{"id":106,"slug":107,"title":108,"created_at":109},"eeac6b9e-ad9d-4831-8eec-8bba3f9bca6a","gap-google-gemini-checkout-fashion-search-zh","Gap 把結帳搬進 Gemini","2026-03-26T07:28:23.937768+00:00",{"id":111,"slug":112,"title":113,"created_at":114},"0740e53f-605d-4d57-8601-c10beb126f3c","google-pushes-gemini-transition-to-march-2026-zh","Google 把 Gemini 轉換延到 2026 年 3…","2026-03-26T07:30:12.825269+00:00",{"id":116,"slug":117,"title":118,"created_at":119},"e660d801-2421-4529-8fa9-86b82b066990","metas-llama-4-benchmark-scandal-gets-worse-zh","Meta Llama 4 分數風波又擴大","2026-03-26T07:34:21.156421+00:00",{"id":121,"slug":122,"title":123,"created_at":124},"183f9e7c-e143-40bb-a6d5-67ba84a3a8bc","accenture-mistral-ai-sovereign-enterprise-deal-zh","Accenture 攜手 Mistral AI 賣主權 AI","2026-03-26T07:38:14.818906+00:00",{"id":126,"slug":127,"title":128,"created_at":129},"191d9b1b-768a-478c-978c-dd7431a38149","mistral-ai-faces-its-hardest-year-yet-zh","Mistral AI 迎來最硬的一年","2026-03-26T07:40:23.716374+00:00"]