[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-workbuddy-harness-engineering-agent-reliability-zh":3,"article-related-workbuddy-harness-engineering-agent-reliability-zh":31,"series-ai-agent-1f24f862-f37b-4bd9-b78d-434713905348":76},{"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":23,"views":27,"created_at":28,"published_at":29,"topic_cluster_id":30},"1f24f862-f37b-4bd9-b78d-434713905348","workbuddy-harness-engineering-agent-reliability-zh","WorkBuddy 證明了 Agent 可靠性不靠大模型本身","\u003Cp data-speakable=\"summary\">200+條可驗證清單才是可用 \u003Ca href=\"\u002Ftag\u002Fagent\">Agent\u003C\u002Fa> 的起點。\u003C\u002Fp>\u003Cp>我站在這一邊：可用的 Agent 不是\u003Ca href=\"\u002Fnews\u002Fgpt-5-6-luna-terra-sol-release-zh\">模型\u003C\u002Fa>能力自然長出來的，而是 Harness 工程把它釘成了產品。\u003C\u002Fp>\u003Cp>WorkBuddy 這篇復盤最有價值的地方，不是再講一遍 \u003Ca href=\"\u002Ftag\u002Fllm\">LLM\u003C\u002Fa>、\u003Ca href=\"\u002Ftag\u002Fmcp\">MCP\u003C\u002Fa> 或 Tool Call，而是把一個現實事實說透了：同樣的模型，進了不同的上下文、權限、驗證和交接機制，表現會像兩個產品。文章裡反覆強調的不是「模型更聰明了」，而是工具\u003Ca href=\"\u002Fnews\u002Fbailian-token-plan-agent-credits-guide-zh\">接入\u003C\u002Fa>、上下文組織、權限邊界、結果驗證、回饋糾正和跨會話延續這些產品側機制，決定了 Agent 能不能在生產環境裡把任務做完。這個判斷很硬，因為它對應的不是抽象理念，而是 WorkBuddy 對國內模型相容、穩定執行和長任務接續的實際效果。\u003C\u002Fp>\u003Ch2>第一個論點\u003C\u002Fh2>\u003Cp>長任務最常見的兩個崩點，文章說得很清楚：一種是 Agent 一次承擔太多工作，跑著跑著上下文耗盡，最後只留下半成品；另一種是它看到部分成果就過早宣佈完成，結果把未驗證的內容當成已完成交付。這不是「模型不夠會想」，而是執行鏈條沒有把任務拆開、驗收和交接做紮實。WorkBuddy 的回應是把任務拆成更細的步驟，把每一步的完成標準寫清楚，再讓 Agent 一次只做一件事。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783987377305-yt32.png\" alt=\"WorkBuddy 證明了 Agent 可靠性不靠大模型本身\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>\u003Ca href=\"\u002Ftag\u002Fanthropic\">Anthropic\u003C\u002Fa> 的 long-running agents 經驗正好提供了對照：初始化 Agent 負責把工作拆成 200+ 條具體行為描述的功能清單，每條都標 pass\u002Ffail，禁止刪條目、降標準；Coding Agent 只負責執行當前項，靠統一啟動腳本、進度文件和 Git 歷史完成跨會話交接、恢復和回滾。這套設計的重點不是「更囉嗦」，而是把長任務從一次性賭命，改成可恢復、可審計、可繼續推進的流水線。沒有這層 Harness，Agent 一旦中途偏航，後面所有聰明都只是在錯誤方向上加速。\u003C\u002Fp>\u003Ch2>第二個論點\u003C\u002Fh2>\u003Cp>文章把模型抽象成無狀態函數，這個前提非常重要。模型本身不記得上一次呼叫，也不會自動知道當前環境、當前檔案、當前權限和當前進度。產品要做的是把系統提示詞、工具、會話歷史、工作區規則、任務狀態和記憶組織成一份「此刻該看到什麼」的上下文，再讓模型基於這份上下文決策。WorkBuddy 不是把所有資訊一股腦塞進去，而是區分寫入、選擇、檢索、壓縮和隔離，避免無關資訊把判斷帶偏。\u003C\u002Fp>\u003Cp>更關鍵的是，權限和高風險動作必須在模型外部攔住。工具呼叫裡，模型只是生成請求，真正持有 \u003Ca href=\"\u002Ftag\u002Fapi\">API\u003C\u002Fa> Key、真正發請求、真正修改資料的是 Agent 執行層，所以審批、參數校驗、Sandbox 和審計日誌不能交給 prompt 口頭約束。文章舉的例子很直接：刪除、發佈、支付這類動作必須確認授權；工具結果過長要截斷並明確告知未完整；錯誤返回不能只給堆疊，還要告訴模型可修正參數和下一步建議。也就是說，Harness 不是包一層皮，而是把「模型會犯錯」當成系統常態來設計。\u003C\u002Fp>\u003Ch2>第三個論點\u003C\u002Fh2>\u003Cp>如果只有 Tool Call，Agent 最多只是一个能點按鈕的模型。WorkBuddy 把 Skill 和 Plugin 拉進來，才真正把「動作」和「流程」分開。Tool 負責一個動作，Skill 負責一類任務的做法，Plugin 則把連接、流程、規則、Hooks 和模板打包成可安裝能力。這個分層很重要，因為現實任務不是「查一次介面就結束」，而是「按倉庫規範讀規則、跑測試、生成說明、處理失敗、再發佈」。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783987376992-p6ae.png\" alt=\"WorkBuddy 證明了 Agent 可靠性不靠大模型本身\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>文章對 Skill 的定義很務實：先讀 AGENTS.md 或 WORKBUDDY.md，再看 git status，識別變更範圍，按需跑格式化、型別檢查和測試，只有驗證通過才生成 PR，發佈前還要確認授權。這個流程的價值在於，它把經驗從「模型記住了一個好習慣」變成「系統強制執行的一套方法」。一旦任務類型變化，Skill 還能按版本更新、評審、回滾。相比之下，把這些流程直接塞進長期記憶，等於讓模型把局部經驗當通用真理，最後只會讓行為越來越飄。\u003C\u002Fp>\u003Ch2>反方可能怎麼說\u003C\u002Fh2>\u003Cp>反對者會說，這套 Harness 設計太重了。Agent 本來就應該讓模型盡量自由地規劃和執行，過多的清單、規則、Skill、權限門檻和驗證步驟，會把系統做成一個臃腫的工作流引擎。尤其在簡單任務上，層層約束會增加延遲、維護成本和上下文占用，甚至讓模型失去靈活性。另一個常見觀點是，模型能力成長很快，今天需要的工程護欄，明天強模型就能自己補上。\u003C\u002Fp>\u003Cp>這個反駁只對一半。簡單任務確實\u003Ca href=\"\u002Fnews\u002Fcrypto-exchanges-should-show-up-in-latam-not-just-advertise-zh\">不該\u003C\u002Fa>上重型 Harness，但生產級 Agent 處理的從來不是單次問答，而是跨工具、跨會話、跨權限邊界的任務。只要任務包含外部狀態、可撤銷操作、長鏈路驗證和多人協作，系統就必須把錯誤關在模型外面。模型變強只能降低出錯概率，不能取消審計、權限、恢復和驗收。WorkBuddy 的價值恰恰在於，它沒有把可靠性押在「模型會自覺」，而是把可靠性寫進了執行結構裡。\u003C\u002Fp>\u003Ch2>你能做什麼\u003C\u002Fh2>\u003Cp>如果你是工程師，就別先追求「讓模型更會想」，先把任務拆分、狀態保存、工具返回、失敗重試、權限審批和驗收標準做成硬約束；如果你是 PM，就把「能用」定義成可恢復、可驗證、可交接，而不是只看一次演示；如果你是 founder，就記住 Agent 產品的競爭點不在 prompt，而在誰能把 Harness 做成可復用、可擴展、可審計的系統能力。\u003C\u002Fp>","WorkBuddy 的案例說明，可用 Agent 的核心不是更強模型，而是更嚴的 Harness 工程。","zhuanlan.zhihu.com","https:\u002F\u002Fzhuanlan.zhihu.com\u002Fp\u002F2059653377424991684",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783987377305-yt32.png","ai-agent","zh","3f5dcc12-bb78-4102-817b-f08f9adbe974",[17,18,19,20,21,22],"WorkBuddy","Agent","Harness工程","可靠性","工具調用","權限控制",[24,25,26],"可用 Agent 的核心不是更強模型，而是更嚴的 Harness。","長任務與高風險動作必須靠拆分、驗收、審計和權限邊界來保證。","Skill 與 Plugin 的價值，在於把經驗變成系統強制執行的流程。",0,"2026-07-14T00:02:31.385926+00:00","2026-07-14T00:02:31.373+00:00","ac421c47-7ea0-4fc6-8634-33f80d04101c",{"tags":32,"relatedLang":35,"relatedPosts":39},[33],{"name":34,"slug":34},"agent",{"id":15,"slug":36,"title":37,"language":38},"workbuddy-harness-engineering-matters-more-than-model-size-en","WorkBuddy Proves Harness Engineering Matters More Than Model Size","en",[40,46,52,58,64,70],{"id":41,"slug":42,"title":43,"cover_image":44,"image_url":44,"created_at":45,"category":13},"415418c7-749e-4352-a21d-d2fa62d8b96b","perplexity-teammate-coding-agent-strategy-zh","Perplexity 應把 Teammate 做成 coding agent，…","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783589576714-n7gs.png","2026-07-09T09:32:25.568287+00:00",{"id":47,"slug":48,"title":49,"cover_image":50,"image_url":50,"created_at":51,"category":13},"884f1bb8-4bae-4cfa-87d1-b323be1d6166","hp-adopts-openai-frontier-global-operations-zh","HP 將 Frontier 送進全球營運","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783362776958-1e1n.png","2026-07-06T18:32:30.897068+00:00",{"id":53,"slug":54,"title":55,"cover_image":56,"image_url":56,"created_at":57,"category":13},"b97b8932-56a9-431f-8270-3f892f8feb94","build-production-vector-db-rag-pipeline-zh","用 n8n 建出可上線的向量資料庫","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783258377362-cg5x.png","2026-07-05T13:32:23.466634+00:00",{"id":59,"slug":60,"title":61,"cover_image":62,"image_url":62,"created_at":63,"category":13},"72b05d2a-7461-4885-a57f-506fd42d714d","ornith-1-agent-coding-server-template-zh","Ornith-1 把代理寫碼變成伺服器","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783085625241-1df8.png","2026-07-03T13:33:20.199747+00:00",{"id":65,"slug":66,"title":67,"cover_image":68,"image_url":68,"created_at":69,"category":13},"662e2729-67d4-4d3c-b42d-ba01e77d5486","crypto-ai-agents-useful-narrow-workflows-zh","Crypto AI 代理有用，但只適合窄流程","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782951466790-q1yp.png","2026-07-02T00:17:19.890956+00:00",{"id":71,"slug":72,"title":73,"cover_image":74,"image_url":74,"created_at":75,"category":13},"04c2d88a-ed27-48ae-8b3c-8cf3bdbc3a5e","ai-agents-in-crypto-2026-protocol-guide-zh","AI 代理幣實作指南","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782950576769-dkjd.png","2026-07-02T00:02:24.596639+00:00",[77,82,87,92,97,102,107,112,117,122],{"id":78,"slug":79,"title":80,"created_at":81},"4ae1e197-1d3d-4233-8733-eafe9cb6438b","claude-now-uses-your-pc-to-finish-tasks-zh","Claude 開始幫你操作電腦","2026-03-26T07:20:48.457387+00:00",{"id":83,"slug":84,"title":85,"created_at":86},"5bede67f-e21c-413d-9ab8-54a3c3d26227","googles-2026-ai-agent-report-decoded-zh","Google 2026 AI Agent 報告解讀","2026-03-26T11:15:22.651956+00:00",{"id":88,"slug":89,"title":90,"created_at":91},"2987d097-563f-46c7-b76f-b558d8ef7c2b","kimi-k25-review-stronger-still-not-legend-zh","Kimi K2.5 評測：更強，但還不是神作","2026-03-27T07:15:55.277513+00:00",{"id":93,"slug":94,"title":95,"created_at":96},"95c9053b-e3f4-4cb5-aace-5c54f4c9e044","claude-code-controls-mac-desktop-zh","Claude Code 也能操控 Mac 了","2026-03-28T03:01:58.58121+00:00",{"id":98,"slug":99,"title":100,"created_at":101},"dc58e153-e3a8-4c06-9b96-1aa64eabbf5f","cloudflare-100x-faster-ai-agent-sandbox-zh","Cloudflare 的 AI 沙箱跑超快","2026-03-28T03:09:44.142236+00:00",{"id":103,"slug":104,"title":105,"created_at":106},"1c8afc56-253f-47a2-979f-1065ff072f2a","openai-backs-isara-agent-swarm-bet-zh","OpenAI 挺 Isara 的 agent swarm …","2026-03-28T03:15:27.513155+00:00",{"id":108,"slug":109,"title":110,"created_at":111},"7379b422-576e-45df-ad5a-d57a0d9dd467","openai-plan-automated-ai-researcher-zh","OpenAI 想做自動化 AI 研究員","2026-03-28T03:17:42.090548+00:00",{"id":113,"slug":114,"title":115,"created_at":116},"48c9889e-86df-450b-a356-e4a4b7c83c5b","harness-engineering-ai-agent-reliability-2026-zh","駕馭工程：從「馬具」到「作業系統」，AI Agent 可靠性的終極密碼","2026-03-31T06:42:53.556721+00:00",{"id":118,"slug":119,"title":120,"created_at":121},"96d8e8c8-1edd-475d-9145-b1e7a1b02b65","mcp-explained-from-prompts-to-production-zh","MCP 怎麼把提示詞變工作流","2026-04-01T09:24:39.321274+00:00",{"id":123,"slug":124,"title":125,"created_at":126},"f2ca7720-b471-4ce5-9336-2a9ac2a876fd","amazon-bedrock-agents-multi-agent-workflows-zh","Amazon Bedrock Agents 進入多代理工作流","2026-04-01T09:30:29.945429+00:00"]