[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-agent-infra-rewrites-ai-infrastructure-zh":3,"article-related-agent-infra-rewrites-ai-infrastructure-zh":28,"series-industry-87497399-e0aa-47bd-b17d-961dd8c683bb":80},{"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":11,"views":25,"created_at":26,"published_at":27,"topic_cluster_id":11},"87497399-e0aa-47bd-b17d-961dd8c683bb","agent-infra-rewrites-ai-infrastructure-zh","Agent 基礎設施正在重寫 AI","\u003Cp>2024 年，\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FSWE-agent\u002FSWE-agent\" target=\"_blank\" rel=\"noopener\">SWE-agent\u003C\u002Fa> 直接把話講白了。Agent 表現，不只看模型本身。工具怎麼接、狀態怎麼存、任務怎麼排，結果差很多。\u003C\u002Fp>\u003Cp>接著，\u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\" target=\"_blank\" rel=\"noopener\">Anthropic\u003C\u002Fa> 又把這件事往前推。它用 \u003Ca href=\"https:\u002F\u002Fmodelcontextprotocol.io\" target=\"_blank\" rel=\"noopener\">Model Context Protocol（MCP）\u003C\u002Fa>，把工具連接變成公開標準。這下子，AI 基礎設施的重心真的開始變了。\u003C\u002Fp>\u003Cp>講白了，以前大家在拼 Tok\u003Ca href=\"\u002Fnews\u002Fopenclaw-security-risks-and-defenses-zh\">en\u003C\u002Fa> 成本。現在更像在拼 Agent 能不能穩穩做事。能不能呼叫工具。能不能記住前一步。能不能在多個系統間協作。\u003C\u002Fp>\u003Ch2>2024 為什麼改變了 Agent 討論\u003C\u002Fh2>\u003Cp>以前的框架很簡單。模型越大，效果越好。參數越多，分數越高。很多人就以為，AI 工程的核心只剩推論服務。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775057778098-qqk6.png\" alt=\"Agent 基礎設施正在重寫 AI\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>但 Agent 出現後，這套說法開始漏氣。只要模型能規劃、能呼叫 API、能讀檔、能重試，周邊基礎設施就立刻變重要。說真的，模型只是大腦。環境才是整個身體。\u003C\u002Fp>\u003Cp>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FSWE-agent\u002FSWE-agent\" target=\"_blank\" rel=\"noopener\">SWE-agent\u003C\u002Fa> 是很好的例子。它把軟體工程任務拿來測，結果很清楚：Agent-computer interface 會直接影響成功率。提示詞格式、工具輸出格式、回饋迴圈，這些都不是小事。\u003C\u002Fp>\u003Cp>Anthropic 的 Agent 指南也在講同一件事。生產環境裡，簡單、可組合的模式通常更穩。這句話很刺耳，但很真。很多團隊還在幻想大框架能解決一切，結果只是把失敗包裝得比較漂亮。\u003C\u002Fp>\u003Cul>\u003Cli>Agent 品質 = 模型能力 + 介面設計。\u003C\u002Fli>\u003Cli>工具呼叫會把問題變成系統協作。\u003C\u002Fli>\u003Cli>越複雜的框架，越容易藏 bug。\u003C\u002Fli>\u003Cli>標準化介面，能少寫很多接線碼。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>AI 基礎設施不再只管 Serving\u003C\u002Fh2>\u003Cp>傳統 AI 基礎設施，重點是 serving。像是 batching、quantization、延遲控制、GPU 利用率。這些還是重要，但不夠了。\u003C\u002Fp>\u003Cp>Agent 會帶來第二層需求。系統要處理工具呼叫、狀態更新、重試、分支流程，甚至多個 worker 同時跑。這就不是單純的推論服務問題了。\u003C\u002Fp>\u003Cp>你可以把它想成三種成本一起冒出來。模型推論成本。協調成本。失敗重試成本。很多團隊只看前者，最後卻被後兩者拖垮。\u003C\u002Fp>\u003Cp>所以現在的 Agent infra，很像 runtime、workflow engine、資料協調層的混血。模型只是其中一塊。真正的產品價值，常常在模型外面。\u003C\u002Fp>\u003Cp>你可能會想問，那實作上要注意什麼？我整理成幾個很實際的點：\u003C\u002Fp>\u003Cul>\u003Cli>Agent 何時要 checkpoint 狀態。\u003C\u002Fli>\u003Cli>哪些動作要同步，哪些可背景執行。\u003C\u002Fli>\u003Cli>工具失敗後，怎麼重試才不會重複寫入。\u003C\u002Fli>\u003Cli>哪些資料能快取，哪些一定要重算。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>MCP 把工具連接往標準化推\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fmodelcontextprotocol.io\" target=\"_blank\" rel=\"noopener\">MCP\u003C\u002Fa> 在 2024 年 11 月公開後，重點不在名字。重點是，它試著把 Agent 和外部工具、外部資料的連接方式標準化。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775057795329-0l9o.png\" alt=\"Agent 基礎設施正在重寫 AI\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這件事很實際。每多一個自訂 connector，就多一種 schema。多一種 auth 流程。多一種 tool description。多一種失敗模式。最後整個系統會變成一坨手工接線。\u003C\u002Fp>\u003Cp>MCP 的價值，就是讓開發者用一致的協定暴露能力。不是每個 app 都要跟每個 m\u003Ca href=\"\u002Fnews\u002Fclaude-code-march-2026-update-fixes-bugs-zh\">ode\u003C\u002Fa>l 單獨對接。這對想保留模型選擇權的團隊，很有感。\u003C\u002Fp>\u003Cp>它也會改變整合成本。內部工具如果能講同一種 protocol，團隊就少花時間改 adapter，多花時間調整 Agent 行為。這種看起來很無聊的基礎工作，往往才是能不能上線的差別。\u003C\u002Fp>\u003Cblockquote>“The future of AI is not about one model to rule them all. It is about many models, many tools, and a standard way for them to talk.” — \u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Fmodel-context-protocol\" target=\"_blank\" rel=\"noopener\">Dario Amodei\u003C\u002Fa>\u003C\u002Fblockquote>\u003Cp>這句話很直白。價值中心，正在從模型存取，移到連接層。\u003C\u002Fp>\u003Ch2>狀態與排程變成第一級問題\u003C\u002Fh2>\u003Cp>Agent 一旦真的開始做事，state 就不再是細節。你要知道它看過什麼。試過什麼。哪些工具輸出可信。任務卡住後要從哪裡接回來。\u003C\u002Fp>\u003Cp>這會把 AI 系統往分散式系統工程拉。不是只有 LLM 問答。還有 checkpoint、recovery、task history、memory pressure。說真的，這些比 prompt engineering 更像正經工程。\u003C\u002Fp>\u003Cp>排程也一樣重要。單一 request 的 Agent，queue 也許就夠了。可是如果你跑上百個長壽命 Agent，就要管 prior\u003Ca href=\"\u002Fnews\u002Fgithub-ai-bug-detection-code-security-zh\">it\u003C\u002Fa>y、concurrency、worker 分配、工具衝突。\u003C\u002Fp>\u003Cp>這時候，Agent infra 通常會拆成四層：\u003C\u002Fp>\u003Cul>\u003Cli>\u003Cstrong>Serving\u003C\u002Fstrong>：推論、batching、延遲、成本。\u003C\u002Fli>\u003Cli>\u003Cstrong>State\u003C\u002Fstrong>：記憶、checkpoint、任務歷史、復原。\u003C\u002Fli>\u003Cli>\u003Cstrong>Scheduling\u003C\u002Fstrong>：佇列、worker、重試、平行執行。\u003C\u002Fli>\u003Cli>\u003Cstrong>Tooling\u003C\u002Fstrong>：connector、權限、schema、protocol。\u003C\u002Fli>\u003C\u002Ful>\u003Cp>這些方向也能從工具生態看出來。\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Flangchain-ai\u002Flanggraph\" target=\"_blank\" rel=\"noopener\">LangGraph\u003C\u002Fa> 主打 graph-based workflow。\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fopenai\u002Fopenai-agents-python\" target=\"_blank\" rel=\"noopener\">OpenAI Agents SDK\u003C\u002Fa> 則偏向結構化多步驟應用。路線不同，但方向很一致。\u003C\u002Fp>\u003Cp>再看數據比較，就更清楚了。傳統 LLM 服務，常常只看 latency 和 cost。Agent 系統還要多看三個指標：工具成功率、狀態恢復率、重試後完成率。少一個，產品就可能卡住。\u003C\u002Fp>\u003Cul>\u003Cli>傳統 serving：看 TPS、P95 latency、GPU 利用率。\u003C\u002Fli>\u003Cli>Agent infra：再加工具成功率、狀態恢復率。\u003C\u002Fli>\u003Cli>單次失敗的代價，常常高於一次推論費用。\u003C\u002Fli>\u003Cli>標準 protocol 會降低整合與維護成本。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>這對現在的建置團隊代表什麼\u003C\u002Fh2>\u003Cp>如果你現在在做 Agent 產品，最直接的建議很簡單。不要再把模型當整個 stack。真正的差距，常常出現在協調失敗有沒有被處理好。\u003C\u002Fp>\u003Cp>工具契約要清楚。狀態要能持久化。流程要能重試。失敗要能復原。這些聽起來很土，但產品能不能穩，通常就卡在這裡。\u003C\u002Fp>\u003Cp>我也想吐槽一下。很多 Agent demo 很會秀。流程拉得很長。抽象層堆得很高。看起來像魔法。可是一進 production，就開始出現各種鬼故事。\u003C\u002Fp>\u003Cp>真正能活下來的系統，通常控制流很清楚。介面很明白。狀態很可追。不是因為它比較酷，是因為它比較不容易炸。\u003C\u002Fp>\u003Ch2>AI 工程正在往哪裡走\u003C\u002Fh2>\u003Cp>這波變化背後，其實是 AI 工程成熟化。早期大家在比模型分數。後來開始比推論成本。現在輪到 Agent 的協調層。\u003C\u002Fp>\u003Cp>這也很像雲端時代的演進。先是 VM。再來是容器。接著是編排。AI 也在走類似路線。模型還是核心，但真正拉開差距的，變成周邊系統設計。\u003C\u002Fp>\u003Cp>所以如果你是工程師、產品人、或創業者，現在該看的不是「哪個模型最強」。而是「哪個 Agent 架構最穩」。差別很大。\u003C\u002Fp>\u003Cp>我自己的判斷很直接。接下來 12 個月，會有更多團隊把重點放在 stateful runtime、protocol support、task orchestration。這些東西會比單純換更大的模型更有感。你如果還在只盯參數量，可能會錯過真正的戰場。\u003C\u002Fp>","SWE-agent、Anthropic 與 MCP 讓人看見，Agent 表現越來越取決於介面、狀態與排程，不再只看模型大小。","zhuanlan.zhihu.com","https:\u002F\u002Fzhuanlan.zhihu.com\u002Fp\u002F2019347592367093406",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775057778098-qqk6.png","industry","zh","d6e69428-99ef-4933-b892-87440d05e2b3",[17,18,19,20,21,22,23,24],"Agent infra","SWE-agent","Anthropic","MCP","AI 基礎設施","LLM","Agent","分散式系統",6,"2026-04-01T10:06:26.981996+00:00","2026-04-01T10:06:26.841+00:00",{"tags":29,"relatedLang":39,"relatedPosts":43},[30,32,34,36,38],{"name":31,"slug":31},"agent",{"name":20,"slug":33},"mcp",{"name":19,"slug":35},"anthropic",{"name":22,"slug":37},"llm",{"name":24,"slug":24},{"id":15,"slug":40,"title":41,"language":42},"agent-infra-rewrites-ai-infrastructure-en","Agent Infra Is Rewriting AI Infrastructure","en",[44,50,56,62,68,74],{"id":45,"slug":46,"title":47,"cover_image":48,"image_url":48,"created_at":49,"category":13},"600a41d7-99a2-48cf-b80e-b28061c65767","andes-technology-20b-risc-v-soc-shipments-zh","Andes RISC-V SoC 出貨破 200 億","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782734588433-8mkx.png","2026-06-29T12:02:32.954092+00:00",{"id":51,"slug":52,"title":53,"cover_image":54,"image_url":54,"created_at":55,"category":13},"383d45a7-2778-436c-902c-fb0d064bfe56","onchain-insurance-proof-institutional-tokenization-test-zh","鏈上保險證明才是機構代幣化的真正考題","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782729171879-ih4l.png","2026-06-29T10:32:25.181256+00:00",{"id":57,"slug":58,"title":59,"cover_image":60,"image_url":60,"created_at":61,"category":13},"e4d40a87-9823-4a96-a9a1-0da241daee68","dtcc-tokenization-link-stellar-zh","DTCC 接上 Stellar，XLM 站上新舞台","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782728271600-sddm.png","2026-06-29T10:17:27.929404+00:00",{"id":63,"slug":64,"title":65,"cover_image":66,"image_url":66,"created_at":67,"category":13},"10f14e61-67c3-4c5e-b561-371efdffb18f","framework-tokenization-ai-financing-fund-zh","Framework 把代幣化變融資","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782727404282-u4vv.png","2026-06-29T10:02:58.99285+00:00",{"id":69,"slug":70,"title":71,"cover_image":72,"image_url":72,"created_at":73,"category":13},"b19bc35b-9d90-4c63-94ab-c46bd759da81","microsoft-investor-relations-page-map-zh","Microsoft 投資人關係頁面地圖","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782726471089-sl2s.png","2026-06-29T09:47:23.941243+00:00",{"id":75,"slug":76,"title":77,"cover_image":78,"image_url":78,"created_at":79,"category":13},"e6695966-a6f4-4b71-ab89-cd61bc205d43","microsoft-190b-ai-capex-plan-msft-452-zh","Microsoft 1900億美元 AI 支出壓力測試 MSFT 452","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782724684761-zzp0.png","2026-06-29T09:17:26.886615+00:00",[81,86,91,96,101,106,111,116,121,126],{"id":82,"slug":83,"title":84,"created_at":85},"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":87,"slug":88,"title":89,"created_at":90},"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":92,"slug":93,"title":94,"created_at":95},"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":97,"slug":98,"title":99,"created_at":100},"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":102,"slug":103,"title":104,"created_at":105},"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":107,"slug":108,"title":109,"created_at":110},"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":112,"slug":113,"title":114,"created_at":115},"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":117,"slug":118,"title":119,"created_at":120},"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":122,"slug":123,"title":124,"created_at":125},"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":127,"slug":128,"title":129,"created_at":130},"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"]