[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-5-ai-agent-papers-worth-tracking-zh":3,"article-related-5-ai-agent-papers-worth-tracking-zh":34,"series-industry-2455fdb3-eaa3-475a-a1e7-1cd98a1c6128":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":25,"views":30,"created_at":31,"published_at":32,"topic_cluster_id":33},"2455fdb3-eaa3-475a-a1e7-1cd98a1c6128","5-ai-agent-papers-worth-tracking-zh","5 個值得追蹤的 AI agent 論文主題","\u003Cp data-speakable=\"summary\">這份整理把 \u003Ca href=\"\u002Ftag\u002Fai-agent\">AI agent\u003C\u002Fa> 論文分成 5 類，讓你快速決定先讀哪個主題。\u003C\u002Fp>\u003Cp>這個由 \u003Ca href=\"https:\u002F\u002Fgithub.com\u002F\">GitHub\u003C\u002Fa> 收錄的論文清單以主題分桶、每兩週更新一次，目前有 1,494 顆星。若你想在不追完整個 arXiv 的情況下，先判斷該從 planning、\u003Ca href=\"\u002Ftag\u002Fskills\">skills\u003C\u002Fa>、harness、survey 還是應用切入，這份清單會直接縮小閱讀範圍。\u003C\u002Fp>\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>項目\u003C\u002Fth>\u003Cth>關注重點\u003C\u002Fth>\u003Cth>典型訊號\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>Harness\u003C\u002Ftd>\u003Ctd>代理執行時的 runtime 結構\u003C\u002Ftd>\u003Ctd>安全、搜尋、工作流\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Skills\u003C\u002Ftd>\u003Ctd>可重複使用的能力模組\u003C\u002Ftd>\u003Ctd>技能生成、治理、評估\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Survey\u003C\u002Ftd>\u003Ctd>領域總覽\u003C\u002Ftd>\u003Ctd>分類法、趨勢、基準\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Architecture\u003C\u002Ftd>\u003Ctd>代理系統如何組織\u003C\u002Ftd>\u003Ctd>單代理、多代理、agent-ops\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Applications\u003C\u002Ftd>\u003Ctd>代理實際落地場景\u003C\u002Ftd>\u003Ctd>Web、軟體、資料、研究\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003Ch2>1. Harness 論文：先看執行層\u003C\u002Fh2>\u003Cp>如果你在意代理怎麼真的跑起來，harness \u003Ca href=\"\u002Fnews\u002Fkucoin-ai-stack-blockchain-plumbing-zh\">區塊\u003C\u002Fa>是最好的起點。它收的是\u003Ca href=\"\u002Fnews\u002Fg7-should-treat-ai-ceos-as-power-brokers-not-guests-zh\">執行\u003C\u002Fa>底座、安全檢查、搜尋行為與架構模式，重點不是提示詞，而是系統在真實流程中的運作方式。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781744565308-wtdy.png\" alt=\"5 個值得追蹤的 AI agent 論文主題\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這一類特別適合做 agent ops、評估或安全設計的人。像 \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.13357\">AI Harness Engineering: A Runtime Substrate for Foundation-Model Software Agents\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.15184\">Is Grep All You Need? How Agent Harnesses Reshape Agentic Search\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.27922v1\">Harness-Bench: Measuring Harness Effects across Models in Realistic Agent Workflows\u003C\u002Fa> 都是在回答同一件事：代理的外層結構會怎麼改變結果。\u003C\u002Fp>\u003Cul>\u003Cli>看的是執行與控制，不只是 prompting\u003C\u002Fli>\u003Cli>適合做 production、eval、safety\u003C\u002Fli>\u003Cli>常和 benchmark、survey 一起出現\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>2. Skills 論文：看可重用能力怎麼長出來\u003C\u002Fh2>\u003Cp>如果你想知道代理能反覆學會什麼，skills 區塊最有操作性。這裡聚焦技能生成、選擇、治理與自我演化，適合把 \u003Ca href=\"\u002Ftag\u002Fagent\">agent\u003C\u002Fa> 視為可組裝系統的人。\u003C\u002Fp>\u003Cp>像 \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.06614\">SkillOS: Learning Skill Curation for Self-Evolving Agents\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.18401\">SkillsVote: Lifecycle Governance of Agent Skills from Collection, Recommendation to Evolution\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.27760\">SkillGrad: Optimizing Agent Skills Like Gradient Descent\u003C\u002Fa> 這些題目，會把技能當成可管理、可評估、可更新的模組來看。\u003C\u002Fp>\u003Ccode>你會在這一區看到的關鍵字：\n- skill generation\n- skill memory and management\n- least-privilege enforcement\n- skill evaluation\n- self-evolving skill systems\u003C\u002Fcode>\u003Ch2>3. Survey 論文：先建立地圖再深入\u003C\u002Fh2>\u003Cp>survey 區塊是最快掌握研究走向的方法。它不押單一方法，而是整理分類、技術路線與未解問題，適合在你還沒決定要鑽哪個子題前先看。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781744565837-8n4g.png\" alt=\"5 個值得追蹤的 AI agent 論文主題\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>像 \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.07358\">A Comprehensive Survey on Agent Skills: Taxonomy, Techniques, and Applications\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.02913\">Generate, Filter, Control, Replay: A Comprehensive Survey of Rollout Strategies for LLM Reinforcement Learning\u003C\u002Fa> 都很符合這個資料庫的風格。它也會延伸到協作、失敗歸因、自我評估等相關工作，適合拿來做文獻綜述的第一層篩選。\u003C\u002Fp>\u003Cul>\u003Cli>適合 literature review 和簡報整理\u003C\u002Fli>\u003Cli>幫你找出值得深讀的子領域\u003C\u002Fli>\u003Cli>搭配 benchmark 類論文最有效\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>4. Architecture 論文：決定系統長相\u003C\u002Fh2>\u003Cp>architecture 區塊把論文按單代理、多代理與 agent-ops 模式整理。這對正在設計產品的人很實用，因為這裡談的不只是模型能力，而是系統怎麼被組成。\u003C\u002Fp>\u003Cp>當你要比較協作方式、分工結構或營運流程時，這一區最有幫助。它也方便你從大方向一路跳到更具體的場景，例如數位代理、企業代理或特定工作流。\u003C\u002Fp>\u003Cul>\u003Cli>單代理適合聚焦任務\u003C\u002Fli>\u003Cli>多代理適合協作與分工\u003C\u002Fli>\u003Cli>agent-ops 和 UX 會影響上線方式\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>5. Applications 論文：直接對準使用場景\u003C\u002Fh2>\u003Cp>applications 區塊是這份 repo 最貼近實作的一段。它把論文分到 embodied、web、mobile、software、data、research、\u003Ca href=\"\u002Ftag\u002Fapi\">API\u003C\u002Fa>、deep research、enterprise、finance 等場景，讀者可以直接挑自己要做的環境。\u003C\u002Fp>\u003Cp>如果你在做瀏覽器代理、寫作助手、程式助理或研究 \u003Ca href=\"\u002Ftag\u002Fcopilot\">copilot\u003C\u002Fa>，這裡能最快縮短清單。你不用先看抽象理論再回頭找案例，而是直接從場景出發反推\u003Ca href=\"\u002Fnews\u002Fping-identity-runtime-identity-ai-agents-zh\">需要\u003C\u002Fa>的能力。\u003C\u002Fp>\u003Ccode>常見的應用群組：\n- Web agents\n- GUI agents\n- Software agents\n- Research agents\n- Enterprise agents\u003C\u002Fcode>\u003Ch2>怎麼挑\u003C\u002Fh2>\u003Cp>如果你在意執行與安全，先看 harness；如果你要的是可重用能力，先看 skills；如果你還在找方向，先讀 survey；如果你已經要做產品，就從 architecture 或 applications 下手。多數讀者最有效的順序是先 survey，再 harness 或 skills，最後進到對應場景。\u003C\u002Fp>\u003Cp>這份 repo 每兩週更新一次，比起一次性的整理，更像會持續長大的閱讀地圖。\u003C\u002Fp>","5 個主題把 AI agent 論文分成 harness、skills、survey、architecture、applications，適合快速決定先讀哪一類。","github.com","https:\u002F\u002Fgithub.com\u002Fmasamasa59\u002Fai-agent-papers",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781744565308-wtdy.png","industry","zh","c54028b3-3c5f-4c00-9c7f-4b8932c74f2d",[17,18,19,20,21,22,23,24],"AI agent","papers","harness","skills","survey","architecture","applications","GitHub repo",[26,27,28,29],"harness 看執行層與安全控制，適合做 agent ops 與評估","skills 把能力模組化，適合關心可重用與自我演化的人","survey 先建立領域地圖，再決定要深讀哪個子題","architecture 與 applications 最適合已經在做產品或場景落地的讀者",0,"2026-06-18T01:02:21.448394+00:00","2026-06-18T01:02:21.442+00:00","caa87b65-9bbc-46fe-bba8-4f4158dd2d8b",{"tags":35,"relatedLang":39,"relatedPosts":43},[36,38],{"name":17,"slug":37},"ai-agent",{"name":20,"slug":20},{"id":15,"slug":40,"title":41,"language":42},"ai-agent-papers-worth-tracking-en","AI agent papers worth tracking in one repo","en",[44,50,56,62,68,74],{"id":45,"slug":46,"title":47,"cover_image":48,"image_url":48,"created_at":49,"category":13},"bd0a5d0d-eb7f-4285-8ee3-680de6bbfb05","90-minute-takedown-turns-ai-ops-into-crisis-zh","90 分鐘下線把 AI 變成事故演練","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781759004216-23ns.png","2026-06-18T05:02:57.07178+00:00",{"id":51,"slug":52,"title":53,"cover_image":54,"image_url":54,"created_at":55,"category":13},"a9ba8f03-c03c-4302-a36d-4ebdb20202f2","gpt-56-fix-and-upgrade-release-zh","GPT-5.6 可能先修再升級，5 個變化先看","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781756273476-5m6b.png","2026-06-18T04:17:27.909445+00:00",{"id":57,"slug":58,"title":59,"cover_image":60,"image_url":60,"created_at":61,"category":13},"e654a80e-57ec-4690-8472-2259f23d0150","anthropic-paid-ai-monetization-path-zh","Anthropic 付費 AI 落地路徑","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781755369678-sep3.png","2026-06-18T04:02:25.602432+00:00",{"id":63,"slug":64,"title":65,"cover_image":66,"image_url":66,"created_at":67,"category":13},"db297e9e-d326-4005-8ca1-487a19c21ca6","github-hottest-repos-ai-agent-tools-zh","GitHub 熱門倉庫都在做 AI agent 工具","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781754478195-my4b.png","2026-06-18T03:47:22.438473+00:00",{"id":69,"slug":70,"title":71,"cover_image":72,"image_url":72,"created_at":73,"category":13},"0700f8ef-d447-41de-bfe4-52991d43746c","anthropic-fable-shows-ai-can-outsmart-constraints-zh","Anthropic Fable 露出 AI 會鑽漏洞","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781751777967-li5i.png","2026-06-18T03:02:33.373632+00:00",{"id":75,"slug":76,"title":77,"cover_image":78,"image_url":78,"created_at":79,"category":13},"8156f591-efd9-45f5-b89e-4f06dcf971dc","openai-partner-network-delivery-strategy-zh","OpenAI 的合作夥伴網路不是 Logo 計畫，而是交付策略","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781741882570-0cji.png","2026-06-18T00:17:18.861629+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"]