{"text": "受Deepseek Harness最近爆火的影響,Codex這邊也放大招了,Codex Harness正式上線。以前我們使用Codex本質上是在使用一個已經做好的智能體,你把任務交給它,它會自己讀代碼、改文件、執行命令、調用工具,最後把事情做完。但這次不一樣,OpenAI開始把Codex背後真正負責讓Agent跑起來的那套系統,也就是Codex Harness。通過Codex SDK和Codex App Server開放出來,你可以把它理解成,以前OpenAI給你的是一台已經裝好的車,而現在連發動機、底盤、控制系統這些核心能力也開始允許開發者直接拿來使用。你不需要重新從零搭一整套Agent,就可以直接基於Codex Harness開發自己的AI數據分析。Agent、辦公Agent、編程Agent,甚至完整的數字員工。Codex Harness和普通模型API最大的區別也就在這裡。普通API的邏輯很簡單,你發一句話,模型返回一個答案,但一個真正能夠完成複雜任務的Agent遠遠不只是問一句答一句。他還需要知道自己現在在哪個項目裡,要讀取哪些文件,當前任務執行到第幾步,某條命令能不能運行,修改文件之前是否需要用戶批准,任務做到一半中斷之後能不能恢復,以及上下文越來越長以後應該怎麼管理。這些真正決定Agent能不能持續穩定完成複雜任務的能力,加在一起就是Harness。換句話說,模型負責思考,Harness負責讓模型真正把事情做完。這一次Codex Harness對開發者最重要的兩個入口,就是Codex SDK和Codex App Server。先說Codex SDK,這個其實最好理解,如果你現在已經有一個Python項目,只是想快速給它增加Codex的Agent能力,就可以直接使用SDK,不需要重新搭一套Agent框架。你可以讓它自動讀取一個代碼倉庫並修改代碼,也可以讓它分析一批數據,然後生成報告,或者執行一個需要連續調用文件、命令行和各種工具的複雜任務。通過SDK,你可以直接啟動一個Codex Thread,也可以繼續之前的任務,恢復歷史任務,甚至Folk出一個新的任務分支。所以Codex SDK更像是一個快速接入入口,你已經有自己的程序,只是希望把Codex Harness的能力加進去,優先用它就可以。而另外一個Codex App Server則更加底層,更適合真正開發完整Agent產品。比如,你想做一個桌面端AI Agent企業內部的數據分析系統,AI編程工具,或者一個帶完整介面的數字員工,這時候僅僅調用SDK就不夠了。因為你的產品不能只是發一句Prompt,然後等最終結果,它還需要時時知道Agent在幹什麼。比如,現在正在執行哪條命令,正在修改哪個文件,要用了哪個工具,任務執行到哪一步,以及某個高權限操作是不是需要用戶確認。Codex App Server就負責把這些過程完整暴露出來。OpenAI官方現在也明確把它作為Codex產品深度集成的重要接口,用來連接Codex Harness和不同的客戶端。這裡可以用一個很簡單的方式理解。Codex Harness是整套Agent執行系統,Codex App Server則是你控制這套系統的接口。你自己的產品不需要重新實現Codex,只需要通過App Server倒數它,開始一個任務,繼續上一次任務,執行下一步,批准這次操作,或者把當前執行狀態返回給前端。所以,如果你只是想快速調用Codex優先使用SDK,如果你想真正開發一款完整的Agent產品,就重點研究App Server。我們團隊也在第一時間做了一次實際測試。之前,我們有一套基於Deep Agents架構開發的AI數據分析。Agent裡面已經包含大量數據分析工具,Skills, MCP,以及一整套完整的數據分析工作流。這一次,我們嘗試把底層Agent Runtime更換為Codex Harness。實際遷移過程比預想中簡單很多。原來的業務邏輯,工具,Skills,和MCP大部分都可以繼續復用。主要變化是把Agent的執行,狀態和任務生命周期管理切換到Codex Harness這一套體系。遷移完成之後,一個比較明顯的感受就是複雜長任務跑起來更加完整和穩定。比如一次完整的數據分析任務,Agent的數據分析任務。Agent可能需要先讀取數據,檢查字段,再寫Python做分析,然後生成圖表,發現異常之後重新計算,最後再生成完整報告。整個過程中可能包含幾十步甚至上百步操作。這種任務真正考驗的,其實已經不只是模型會不會寫代碼,而是Agent在執行幾十步以後會不會忘記最開始的目標?會不會執行一半停下來?報錯之後能不能繼續?以及任務中斷之後能不能恢復?這些問題恰恰就是Harness最核心的價值。Codex Harness本身也已經把很多過去開發Agent最麻煩的事情提前處理好了。比如Sandbox可以把Agent的命令執行,限制在安全環境裡。Approval可以讓敏感操作先經過用戶確認。Thread可以把一個完整任務和狀態保存下來,下次繼續執行。同時還有MCP和Skills可以繼續給Agent增加外部工具和專業能力。如果這些能力全部從0自己開發,成本其實非常高。而現在Codex Harness已經把它們變成了一套可以直接附用的基礎設施。這也是為什麼我覺得這次Codex Harness真正重要的地方,並不是OpenAI又發布了一個新的Agent SDK,而是開發Agent的方式正在發生變化。以前我們做Agent往往是先選一個模型,再選LangChain,LangGraph或者其他框架,然後自己搭工具,狀態,全線,memory和工作流。以後則很可能變成先選擇一套成熟的Harness,然後直接在這套Harness上做自己的業務。模型可以換,業務也可以換,但底層那套讓Agent穩定執行任務的系統可以直接附用。而從OpenAI目前公開的信息來看,Codex CLI,Codex App,IDE等不同形態,本身就是圍繞同一套Codex Harness能力展開的。現在真正發生變化的是,這套原本主要服務OpenAI自己Codex產品的能力,正在進一步向開發者開放。所以Codex的定位也開始不再只是一個AI變成工具,它正在逐漸變成一套可以拿來開發其他智能體的通用Agent基座。以後你想開發AI數據分析系統,研發Agent,辦公Agent,甚至企業內部數字員工,並不一定還要從最底層重新打一整套Agent Framework,完全可以直接站在Codex Harness已經搭好的基礎設施上,繼續做自己的產品和業務能力。本期後面的實戰,我們就會直接把Codex Harness跑起來,分別演示Codex SDK和Codex App Server的使用方式,以及如何接入工具,Skills和MCP。最後再把它真正集成進一套完整的AI數據分析Agent。同時,我們團隊也已經整理好了Codex Harness使用與開發教程,完整配置說明,以及這套基於Codex Harness重構的AI數據分析Agent項目緣碼。想直接跟著項目復線的,可以查看制定評論領取。如果之前大家學習Codex,重點還是怎麼把Codex用好,那麼從Codex Harness開始,一個更值得關注的問題已經變成了,怎麼拿Codex做出屬於自己的Agent。Codex Harness", "segments": [{"id": 0, "seek": 0, "start": 0.0, "end": 5.82, "text": "受Deepseek Harness最近爆火的影響,Codex這邊也放大招了,Codex Harness正式上線。", "tokens": [50365, 23151, 11089, 595, 405, 916, 389, 16937, 46777, 45827, 26171, 1546, 16820, 13665, 123, 11, 34, 1429, 87, 22821, 6404, 12744, 3582, 6852, 249, 2289, 11, 34, 1429, 87, 389, 16937, 15789, 27584, 5708, 31699, 1543, 50656], "temperature": 0, "avg_logprob": -0.13710655578195233, "compression_ratio": 1.1891891891891893, "no_speech_prob": 7.632743395658004e-12}, {"id": 1, "seek": 0, "start": 5.82, "end": 17.7, "text": "以前我們使用Codex本質上是在使用一個已經做好的智能體,你把任務交給它,它會自己讀代碼、改文件、執行命令、調用工具,最後把事情做完。", "tokens": [50656, 44996, 5884, 22982, 9254, 34, 1429, 87, 8802, 43450, 5708, 1541, 3581, 22982, 9254, 8990, 25338, 10907, 20715, 5094, 118, 8225, 23987, 11, 2166, 16075, 26443, 39517, 28455, 17798, 11284, 11, 11284, 6236, 17645, 7422, 222, 19105, 16337, 120, 1231, 34490, 17174, 20485, 1231, 16262, 115, 8082, 25236, 49061, 1231, 32943, 9254, 23323, 39806, 11, 20578, 16075, 24675, 10907, 14128, 1543, 51250], "temperature": 0, "avg_logprob": -0.13710655578195233, "compression_ratio": 1.1891891891891893, "no_speech_prob": 7.632743395658004e-12}, {"id": 2, "seek": 0, "start": 17.7, "end": 25.6, "text": "但這次不一樣,OpenAI開始把Codex背後真正負責讓Agent跑起來的那套系統,也就是Codex Harness。", "tokens": [51250, 8395, 2664, 9487, 1960, 22977, 11, 45569, 48698, 21017, 16075, 34, 1429, 87, 46329, 5661, 6303, 15789, 11561, 254, 11561, 105, 21195, 32, 6930, 32585, 21670, 1546, 4184, 1881, 245, 25368, 33725, 11, 6404, 5620, 34, 1429, 87, 389, 16937, 1543, 51645], "temperature": 0, "avg_logprob": -0.13710655578195233, "compression_ratio": 1.1891891891891893, "no_speech_prob": 7.632743395658004e-12}, {"id": 3, "seek": 2560, "start": 25.6, "end": 42.2, "text": "通過Codex SDK和Codex App Server開放出來,你可以把它理解成,以前OpenAI給你的是一台已經裝好的車,而現在連發動機、底盤、控制系統這些核心能力也開始允許開發者直接拿來使用。", "tokens": [50365, 19550, 8816, 34, 1429, 87, 37135, 12565, 34, 1429, 87, 3132, 25684, 8949, 12744, 29741, 11, 42766, 42061, 13876, 17278, 11336, 11, 44996, 45569, 48698, 17798, 18961, 1541, 2257, 15433, 25338, 44453, 20715, 25941, 11, 11070, 12648, 36252, 14637, 12572, 17543, 1231, 22816, 5419, 97, 1231, 48707, 25491, 25368, 33725, 29869, 8987, 116, 7945, 8225, 13486, 6404, 21017, 2347, 223, 40914, 8949, 14637, 12444, 43297, 24351, 3763, 22982, 9254, 1543, 51195], "temperature": 0, "avg_logprob": -0.09697803661977644, "compression_ratio": 1.172514619883041, "no_speech_prob": 1.0785574690308497e-11}, {"id": 4, "seek": 2560, "start": 42.2, "end": 49.0, "text": "你不需要重新從零搭一整套Agent,就可以直接基於Codex Harness開發自己的AI數據分析。", "tokens": [51195, 2166, 1960, 35748, 12624, 12560, 21068, 6306, 114, 22220, 255, 2257, 27662, 1881, 245, 32, 6930, 11, 36781, 43297, 26008, 19488, 34, 1429, 87, 389, 16937, 8949, 14637, 17645, 1546, 48698, 30622, 36841, 6627, 7360, 238, 1543, 51535], "temperature": 0, "avg_logprob": -0.09697803661977644, "compression_ratio": 1.172514619883041, "no_speech_prob": 1.0785574690308497e-11}, {"id": 5, "seek": 2560, "start": 49.0, "end": 53.2, "text": "Agent、辦公Agent、編程Agent,甚至完整的數字員工。", "tokens": [51535, 32, 6930, 1231, 23626, 13545, 32, 6930, 1231, 12608, 101, 29649, 32, 6930, 11, 17556, 20844, 14128, 27662, 1546, 30622, 22381, 20992, 23323, 1543, 51745], "temperature": 0, "avg_logprob": -0.09697803661977644, "compression_ratio": 1.172514619883041, "no_speech_prob": 1.0785574690308497e-11}, {"id": 6, "seek": 5320, "start": 53.2, "end": 58.300000000000004, "text": "Codex Harness和普通模型API最大的區別也就在這裡。", "tokens": [50365, 34, 1429, 87, 389, 16937, 12565, 29993, 19550, 41908, 39823, 4715, 40, 8661, 39156, 44526, 16158, 6404, 3111, 3581, 18771, 1543, 50620], "temperature": 0, "avg_logprob": -0.1177926103607947, "compression_ratio": 1.164835164835165, "no_speech_prob": 1.5394583177674726e-11}, {"id": 7, "seek": 5320, "start": 58.300000000000004, "end": 68.5, "text": "普通API的邏輯很簡單,你發一句話,模型返回一個答案,但一個真正能夠完成複雜任務的Agent遠遠不只是問一句答一句。", "tokens": [50620, 29993, 19550, 4715, 40, 1546, 3023, 237, 12290, 107, 4563, 31995, 11, 2166, 14637, 2257, 34592, 11103, 11, 41908, 39823, 3316, 242, 8350, 8990, 28223, 28899, 11, 8395, 8990, 6303, 15789, 8225, 31649, 41509, 164, 97, 229, 6306, 250, 26443, 39517, 1546, 32, 6930, 34489, 34489, 1960, 36859, 11361, 2257, 34592, 28223, 2257, 34592, 1543, 51130], "temperature": 0, "avg_logprob": -0.1177926103607947, "compression_ratio": 1.164835164835165, "no_speech_prob": 1.5394583177674726e-11}, {"id": 8, "seek": 5320, "start": 68.5, "end": 74.2, "text": "他還需要知道自己現在在哪個項目裡,要讀取哪些文件,當前任務執行到第幾步,", "tokens": [51130, 5000, 7824, 35748, 7758, 17645, 12648, 3581, 17028, 3338, 8313, 227, 11386, 11066, 11, 4275, 7422, 222, 29436, 17028, 13824, 17174, 20485, 11, 13118, 8945, 26443, 39517, 16262, 115, 8082, 4511, 7536, 23575, 31429, 11, 51415], "temperature": 0, "avg_logprob": -0.1177926103607947, "compression_ratio": 1.164835164835165, "no_speech_prob": 1.5394583177674726e-11}, {"id": 9, "seek": 7420, "start": 74.2, "end": 82.3, "text": "某條命令能不能運行,修改文件之前是否需要用戶批准,任務做到一半中斷之後能不能恢復,", "tokens": [50365, 17238, 238, 31106, 25236, 49061, 8225, 28590, 32772, 8082, 11, 7792, 106, 34490, 17174, 20485, 32442, 1541, 45334, 35748, 9254, 1486, 114, 3416, 117, 6336, 228, 11, 26443, 39517, 10907, 4511, 2257, 30018, 5975, 4307, 115, 26362, 8225, 28590, 11845, 95, 2172, 102, 11, 50770], "temperature": 0, "avg_logprob": -0.14226011129525992, "compression_ratio": 1.0961538461538463, "no_speech_prob": 1.579293987252761e-11}, {"id": 10, "seek": 7420, "start": 82.3, "end": 85.3, "text": "以及上下文越來越長以後應該怎麼管理。", "tokens": [50770, 40282, 5708, 4438, 17174, 25761, 3763, 25761, 15353, 3588, 5661, 26087, 11614, 23131, 13876, 1543, 50920], "temperature": 0, "avg_logprob": -0.14226011129525992, "compression_ratio": 1.0961538461538463, "no_speech_prob": 1.579293987252761e-11}, {"id": 11, "seek": 8530, "start": 85.3, "end": 93.39999999999999, "text": "這些真正決定Agent能不能持續穩定完成複雜任務的能力,加在一起就是Harness。", "tokens": [50365, 29869, 6303, 15789, 33540, 12088, 32, 6930, 8225, 28590, 17694, 29583, 13740, 102, 12088, 41509, 164, 97, 229, 6306, 250, 26443, 39517, 1546, 8225, 13486, 11, 9990, 3581, 29567, 5620, 39, 16937, 1543, 50770], "temperature": 0, "avg_logprob": -0.06184751400049182, "compression_ratio": 1.0595238095238095, "no_speech_prob": 1.6397255081512796e-11}, {"id": 12, "seek": 8530, "start": 93.39999999999999, "end": 98.39999999999999, "text": "換句話說,模型負責思考,Harness負責讓模型真正把事情做完。", "tokens": [50770, 36338, 34592, 11103, 4622, 11, 41908, 39823, 11561, 254, 11561, 105, 8870, 26504, 11, 39, 16937, 11561, 254, 11561, 105, 21195, 41908, 39823, 6303, 15789, 16075, 24675, 10907, 14128, 1543, 51020], "temperature": 0, "avg_logprob": -0.06184751400049182, "compression_ratio": 1.0595238095238095, "no_speech_prob": 1.6397255081512796e-11}, {"id": 13, "seek": 9840, "start": 98.4, "end": 106.5, "text": "這一次Codex Harness對開發者最重要的兩個入口,就是Codex SDK和Codex App Server。", "tokens": [50365, 2664, 27505, 34, 1429, 87, 389, 16937, 2855, 8949, 14637, 12444, 8661, 24928, 1546, 34623, 14028, 18144, 11, 5620, 34, 1429, 87, 37135, 12565, 34, 1429, 87, 3132, 25684, 1543, 50770], "temperature": 0, "avg_logprob": -0.10990549723307291, "compression_ratio": 1.0984848484848484, "no_speech_prob": 1.6973873695480535e-11}, {"id": 14, "seek": 9840, "start": 106.5, "end": 115.5, "text": "先說Codex SDK,這個其實最好理解,如果你現在已經有一個Python項目,只是想快速給它增加Codex的Agent能力,", "tokens": [50770, 10108, 4622, 34, 1429, 87, 37135, 11, 6287, 14139, 8661, 2131, 13876, 17278, 11, 45669, 12648, 25338, 2412, 8990, 47, 88, 11943, 8313, 227, 11386, 11, 36859, 7093, 10251, 31217, 17798, 11284, 24228, 252, 9990, 34, 1429, 87, 1546, 32, 6930, 8225, 13486, 11, 51220], "temperature": 0, "avg_logprob": -0.10990549723307291, "compression_ratio": 1.0984848484848484, "no_speech_prob": 1.6973873695480535e-11}, {"id": 15, "seek": 9840, "start": 115.5, "end": 120.5, "text": "就可以直接使用SDK,不需要重新搭一套Agent框架。", "tokens": [51220, 36781, 43297, 22982, 9254, 23969, 42, 11, 1960, 35748, 12624, 12560, 22220, 255, 2257, 1881, 245, 32, 6930, 21416, 228, 7360, 114, 1543, 51470], "temperature": 0, "avg_logprob": -0.10990549723307291, "compression_ratio": 1.0984848484848484, "no_speech_prob": 1.6973873695480535e-11}, {"id": 16, "seek": 12050, "start": 120.5, "end": 132.1, "text": "你可以讓它自動讀取一個代碼倉庫並修改代碼,也可以讓它分析一批數據,然後生成報告,或者執行一個需要連續調用文件、命令行和各種工具的複雜任務。", "tokens": [50365, 42766, 21195, 11284, 9722, 12572, 7422, 222, 29436, 8990, 19105, 16337, 120, 2122, 231, 6346, 104, 35934, 7792, 106, 34490, 19105, 16337, 120, 11, 47935, 21195, 11284, 6627, 7360, 238, 2257, 3416, 117, 30622, 36841, 11, 10213, 8244, 11336, 17803, 16846, 11, 31148, 16262, 115, 8082, 8990, 35748, 36252, 29583, 32943, 9254, 17174, 20485, 1231, 25236, 49061, 8082, 12565, 17516, 16516, 23323, 39806, 1546, 164, 97, 229, 6306, 250, 26443, 39517, 1543, 50945], "temperature": 0, "avg_logprob": -0.07982709861937023, "compression_ratio": 1.2952646239554317, "no_speech_prob": 1.533191282265811e-11}, {"id": 17, "seek": 12050, "start": 132.1, "end": 141.6, "text": "通過SDK,你可以直接啟動一個Codex Thread,也可以繼續之前的任務,恢復歷史任務,甚至Folk出一個新的任務分支。", "tokens": [50945, 19550, 8816, 23969, 42, 11, 42766, 43297, 3284, 253, 12572, 8990, 34, 1429, 87, 334, 2538, 11, 47935, 38459, 32442, 1546, 26443, 39517, 11, 11845, 95, 2172, 102, 4287, 115, 45399, 26443, 39517, 11, 17556, 20844, 37, 401, 74, 7781, 8990, 12560, 1546, 26443, 39517, 6627, 24400, 1543, 51420], "temperature": 0, "avg_logprob": -0.07982709861937023, "compression_ratio": 1.2952646239554317, "no_speech_prob": 1.533191282265811e-11}, {"id": 18, "seek": 12050, "start": 141.6, "end": 150.3, "text": "所以Codex SDK更像是一個快速接入入口,你已經有自己的程序,只是希望把Codex Harness的能力加進去,", "tokens": [51420, 7239, 34, 1429, 87, 37135, 19002, 12760, 1541, 8990, 10251, 31217, 14468, 14028, 14028, 18144, 11, 2166, 25338, 2412, 17645, 1546, 29649, 6346, 237, 11, 36859, 29955, 16075, 34, 1429, 87, 389, 16937, 1546, 8225, 13486, 9990, 18214, 6734, 11, 51855], "temperature": 0, "avg_logprob": -0.07982709861937023, "compression_ratio": 1.2952646239554317, "no_speech_prob": 1.533191282265811e-11}, {"id": 19, "seek": 15030, "start": 150.3, "end": 157.9, "text": "優先用它就可以。而另外一個Codex App Server則更加底層,更適合真正開發完整Agent產品。", "tokens": [50365, 12827, 103, 10108, 9254, 11284, 36781, 1543, 11070, 26202, 8990, 34, 1429, 87, 3132, 25684, 46225, 19002, 9990, 22816, 9636, 97, 11, 19002, 3330, 102, 14245, 6303, 15789, 8949, 14637, 14128, 27662, 32, 6930, 33299, 30246, 1543, 50745], "temperature": 0, "avg_logprob": -0.1187402488541429, "compression_ratio": 1.1831831831831832, "no_speech_prob": 1.722161128980826e-11}, {"id": 20, "seek": 15030, "start": 157.9, "end": 169.4, "text": "比如,你想做一個桌面端AI Agent企業內部的數據分析系統,AI編程工具,或者一個帶完整介面的數字員工,這時候僅僅調用SDK就不夠了。", "tokens": [50745, 36757, 11, 38386, 10907, 8990, 162, 5475, 8833, 11957, 107, 48698, 27174, 7384, 223, 27119, 28472, 13470, 1546, 30622, 36841, 6627, 7360, 238, 25368, 33725, 11, 48698, 12608, 101, 29649, 23323, 39806, 11, 31148, 8990, 32379, 14128, 27662, 30312, 8833, 1546, 30622, 22381, 20992, 23323, 11, 2664, 14010, 9502, 227, 9502, 227, 32943, 9254, 23969, 42, 3111, 1960, 31649, 2289, 1543, 51320], "temperature": 0, "avg_logprob": -0.1187402488541429, "compression_ratio": 1.1831831831831832, "no_speech_prob": 1.722161128980826e-11}, {"id": 21, "seek": 15030, "start": 169.4, "end": 175.60000000000002, "text": "因為你的產品不能只是發一句Prompt,然後等最終結果,它還需要時時知道Agent在幹什麼。", "tokens": [51320, 11471, 18961, 33299, 30246, 28590, 36859, 14637, 2257, 34592, 47, 4397, 662, 11, 10213, 10187, 8661, 29371, 35181, 11, 11284, 7824, 35748, 6611, 6611, 7758, 32, 6930, 3581, 29932, 7598, 1543, 51630], "temperature": 0, "avg_logprob": -0.1187402488541429, "compression_ratio": 1.1831831831831832, "no_speech_prob": 1.722161128980826e-11}, {"id": 22, "seek": 17560, "start": 175.6, "end": 183.2, "text": "比如,現在正在執行哪條命令,正在修改哪個文件,要用了哪個工具,任務執行到哪一步,", "tokens": [50365, 36757, 11, 12648, 15789, 3581, 16262, 115, 8082, 17028, 31106, 25236, 49061, 11, 15789, 3581, 7792, 106, 34490, 17028, 3338, 17174, 20485, 11, 4275, 9254, 2289, 17028, 3338, 23323, 39806, 11, 26443, 39517, 16262, 115, 8082, 4511, 17028, 2257, 31429, 11, 50745], "temperature": 0, "avg_logprob": -0.06757910323865486, "compression_ratio": 1.1872909698996656, "no_speech_prob": 1.6084123616311174e-11}, {"id": 23, "seek": 17560, "start": 183.2, "end": 192.0, "text": "以及某個高權限操作是不是需要用戶確認。Codex App Server就負責把這些過程完整暴露出來。", "tokens": [50745, 40282, 17238, 238, 3338, 12979, 40212, 43446, 13167, 235, 11914, 23034, 35748, 9254, 1486, 114, 24293, 22041, 1543, 34, 1429, 87, 3132, 25684, 3111, 11561, 254, 11561, 105, 16075, 29869, 8816, 29649, 14128, 27662, 23826, 112, 18594, 110, 29741, 1543, 51185], "temperature": 0, "avg_logprob": -0.06757910323865486, "compression_ratio": 1.1872909698996656, "no_speech_prob": 1.6084123616311174e-11}, {"id": 24, "seek": 17560, "start": 192.0, "end": 201.2, "text": "OpenAI官方現在也明確把它作為Codex產品深度集成的重要接口,用來連接Codex Harness和不同的客戶端。", "tokens": [51185, 45569, 48698, 31929, 9249, 12648, 6404, 11100, 24293, 42061, 11914, 6344, 34, 1429, 87, 33299, 30246, 24043, 13127, 26020, 11336, 1546, 24928, 14468, 18144, 11, 9254, 3763, 36252, 14468, 34, 1429, 87, 389, 16937, 12565, 47123, 1546, 32316, 1486, 114, 11957, 107, 1543, 51645], "temperature": 0, "avg_logprob": -0.06757910323865486, "compression_ratio": 1.1872909698996656, "no_speech_prob": 1.6084123616311174e-11}, {"id": 25, "seek": 20120, "start": 201.2, "end": 210.29999999999998, "text": "這裡可以用一個很簡單的方式理解。Codex Harness是整套Agent執行系統,Codex App Server則是你控制這套系統的接口。", "tokens": [50365, 18771, 6723, 9254, 8990, 4563, 31995, 1546, 9249, 27584, 13876, 17278, 1543, 34, 1429, 87, 389, 16937, 1541, 27662, 1881, 245, 32, 6930, 16262, 115, 8082, 25368, 33725, 11, 34, 1429, 87, 3132, 25684, 46225, 1541, 2166, 48707, 25491, 2664, 1881, 245, 25368, 33725, 1546, 14468, 18144, 1543, 50820], "temperature": 0, "avg_logprob": -0.06979044022098664, "compression_ratio": 1.215547703180212, "no_speech_prob": 1.7803550300676818e-11}, {"id": 26, "seek": 20120, "start": 210.29999999999998, "end": 223.6, "text": "你自己的產品不需要重新實現Codex,只需要通過App Server倒數它,開始一個任務,繼續上一次任務,執行下一步,批准這次操作,或者把當前執行狀態返回給前端。", "tokens": [50820, 2166, 17645, 1546, 33299, 30246, 1960, 35748, 12624, 12560, 10376, 9581, 34, 1429, 87, 11, 14003, 35748, 19550, 8816, 9132, 25684, 30878, 30622, 11284, 11, 21017, 8990, 26443, 39517, 11, 38459, 5708, 27505, 26443, 39517, 11, 16262, 115, 8082, 45271, 31429, 11, 3416, 117, 6336, 228, 2664, 9487, 13167, 235, 11914, 11, 31148, 16075, 13118, 8945, 16262, 115, 8082, 18637, 222, 45248, 3316, 242, 8350, 17798, 8945, 11957, 107, 1543, 51485], "temperature": 0, "avg_logprob": -0.06979044022098664, "compression_ratio": 1.215547703180212, "no_speech_prob": 1.7803550300676818e-11}, {"id": 27, "seek": 22360, "start": 223.6, "end": 232.7, "text": "所以,如果你只是想快速調用Codex優先使用SDK,如果你想真正開發一款完整的Agent產品,就重點研究App Server。", "tokens": [50365, 7239, 11, 45669, 36859, 7093, 10251, 31217, 32943, 9254, 34, 1429, 87, 12827, 103, 10108, 22982, 9254, 23969, 42, 11, 13119, 38386, 6303, 15789, 8949, 14637, 2257, 48798, 14128, 27662, 1546, 32, 6930, 33299, 30246, 11, 3111, 12624, 8216, 23230, 242, 44704, 9132, 25684, 1543, 50820], "temperature": 0, "avg_logprob": -0.12147259712219238, "compression_ratio": 1.2110389610389611, "no_speech_prob": 1.5446368142879585e-11}, {"id": 28, "seek": 22360, "start": 232.7, "end": 235.7, "text": "我們團隊也在第一時間做了一次實際測試。", "tokens": [50820, 5884, 37766, 33929, 6404, 3581, 18049, 20788, 10907, 2289, 27505, 10376, 34837, 9592, 105, 22099, 1543, 50970], "temperature": 0, "avg_logprob": -0.12147259712219238, "compression_ratio": 1.2110389610389611, "no_speech_prob": 1.5446368142879585e-11}, {"id": 29, "seek": 22360, "start": 235.7, "end": 240.7, "text": "之前,我們有一套基於Deep Agents架構開發的AI數據分析。", "tokens": [50970, 32442, 11, 5884, 32241, 1881, 245, 26008, 19488, 11089, 595, 2725, 791, 7360, 114, 43362, 8949, 14637, 1546, 48698, 30622, 36841, 6627, 7360, 238, 1543, 51220], "temperature": 0, "avg_logprob": -0.12147259712219238, "compression_ratio": 1.2110389610389611, "no_speech_prob": 1.5446368142879585e-11}, {"id": 30, "seek": 22360, "start": 240.7, "end": 246.7, "text": "Agent裡面已經包含大量數據分析工具,Skills, MCP,以及一整套完整的數據分析工作流。", "tokens": [51220, 32, 6930, 32399, 25338, 23305, 2392, 104, 3582, 26748, 30622, 36841, 6627, 7360, 238, 23323, 39806, 11, 50, 74, 2565, 11, 8797, 47, 11, 40282, 2257, 27662, 1881, 245, 14128, 27662, 1546, 30622, 36841, 6627, 7360, 238, 41315, 27854, 1543, 51520], "temperature": 0, "avg_logprob": -0.12147259712219238, "compression_ratio": 1.2110389610389611, "no_speech_prob": 1.5446368142879585e-11}, {"id": 31, "seek": 24670, "start": 246.7, "end": 251.79999999999998, "text": "這一次,我們嘗試把底層Agent Runtime更換為Codex Harness。", "tokens": [50365, 2664, 27505, 11, 5884, 10756, 245, 22099, 16075, 22816, 9636, 97, 32, 6930, 497, 2760, 1312, 19002, 36338, 6344, 34, 1429, 87, 389, 16937, 1543, 50620], "temperature": 0, "avg_logprob": -0.16360592312282987, "compression_ratio": 1.2712328767123289, "no_speech_prob": 1.474336451534608e-11}, {"id": 32, "seek": 24670, "start": 251.79999999999998, "end": 254.79999999999998, "text": "實際遷移過程比預想中簡單很多。", "tokens": [50620, 10376, 34837, 3330, 115, 8204, 119, 8816, 29649, 11706, 30156, 7093, 5975, 31995, 20778, 1543, 50770], "temperature": 0, "avg_logprob": -0.16360592312282987, "compression_ratio": 1.2712328767123289, "no_speech_prob": 1.474336451534608e-11}, {"id": 33, "seek": 24670, "start": 254.79999999999998, "end": 259.8, "text": "原來的業務邏輯,工具,Skills,和MCP大部分都可以繼續復用。", "tokens": [50770, 19683, 3763, 1546, 27119, 39517, 3023, 237, 12290, 107, 11, 23323, 39806, 11, 50, 74, 2565, 11, 12565, 44, 20049, 3582, 32174, 7182, 6723, 38459, 2172, 102, 9254, 1543, 51020], "temperature": 0, "avg_logprob": -0.16360592312282987, "compression_ratio": 1.2712328767123289, "no_speech_prob": 1.474336451534608e-11}, {"id": 34, "seek": 24670, "start": 259.8, "end": 266.8, "text": "主要變化是把Agent的執行,狀態和任務生命周期管理切換到Codex Harness這一套體系。", "tokens": [51020, 13557, 4275, 26327, 23756, 1541, 16075, 32, 6930, 1546, 16262, 115, 8082, 11, 18637, 222, 45248, 12565, 26443, 39517, 8244, 25236, 37390, 16786, 23131, 13876, 23632, 36338, 4511, 34, 1429, 87, 389, 16937, 32260, 1881, 245, 23987, 25368, 1543, 51370], "temperature": 0, "avg_logprob": -0.16360592312282987, "compression_ratio": 1.2712328767123289, "no_speech_prob": 1.474336451534608e-11}, {"id": 35, "seek": 24670, "start": 266.8, "end": 272.8, "text": "遷移完成之後,一個比較明顯的感受就是複雜長任務跑起來更加完整和穩定。", "tokens": [51370, 3330, 115, 8204, 119, 41509, 26362, 11, 8990, 25174, 11100, 46605, 1546, 9709, 23151, 5620, 164, 97, 229, 6306, 250, 15353, 26443, 39517, 32585, 21670, 19002, 9990, 14128, 27662, 12565, 13740, 102, 12088, 1543, 51670], "temperature": 0, "avg_logprob": -0.16360592312282987, "compression_ratio": 1.2712328767123289, "no_speech_prob": 1.474336451534608e-11}, {"id": 36, "seek": 24670, "start": 272.8, "end": 275.8, "text": "比如一次完整的數據分析任務,Agent的數據分析任務。", "tokens": [51670, 36757, 27505, 14128, 27662, 1546, 30622, 36841, 6627, 7360, 238, 26443, 39517, 11, 32, 6930, 1546, 30622, 36841, 6627, 7360, 238, 26443, 39517, 1543, 51820], "temperature": 0, "avg_logprob": -0.16360592312282987, "compression_ratio": 1.2712328767123289, "no_speech_prob": 1.474336451534608e-11}, {"id": 37, "seek": 27580, "start": 275.8, "end": 285.8, "text": "Agent可能需要先讀取數據,檢查字段,再寫Python做分析,然後生成圖表,發現異常之後重新計算,最後再生成完整報告。", "tokens": [50365, 32, 6930, 16657, 35748, 10108, 7422, 222, 29436, 30622, 36841, 11, 40741, 95, 42623, 22381, 28427, 11, 8623, 4510, 104, 47, 88, 11943, 10907, 6627, 7360, 238, 11, 10213, 8244, 11336, 2523, 244, 17571, 11, 38927, 6904, 108, 11279, 26362, 12624, 12560, 30114, 19497, 11, 20578, 8623, 8244, 11336, 14128, 27662, 17803, 16846, 1543, 50865], "temperature": 0, "avg_logprob": -0.09867606707067297, "compression_ratio": 1.2615384615384615, "no_speech_prob": 1.733909023304836e-11}, {"id": 38, "seek": 27580, "start": 285.8, "end": 289.8, "text": "整個過程中可能包含幾十步甚至上百步操作。", "tokens": [50865, 27662, 3338, 8816, 29649, 5975, 16657, 23305, 2392, 104, 23575, 20145, 31429, 17556, 20844, 5708, 31906, 31429, 13167, 235, 11914, 1543, 51065], "temperature": 0, "avg_logprob": -0.09867606707067297, "compression_ratio": 1.2615384615384615, "no_speech_prob": 1.733909023304836e-11}, {"id": 39, "seek": 27580, "start": 289.8, "end": 299.8, "text": "這種任務真正考驗的,其實已經不只是模型會不會寫代碼,而是Agent在執行幾十步以後會不會忘記最開始的目標?", "tokens": [51065, 33005, 26443, 39517, 6303, 15789, 26504, 24023, 245, 1546, 11, 14139, 25338, 1960, 36859, 41908, 39823, 6236, 21121, 4510, 104, 19105, 16337, 120, 11, 11070, 1541, 32, 6930, 3581, 16262, 115, 8082, 23575, 20145, 31429, 3588, 5661, 6236, 21121, 26677, 16958, 8661, 21017, 1546, 11386, 46299, 30, 51565], "temperature": 0, "avg_logprob": -0.09867606707067297, "compression_ratio": 1.2615384615384615, "no_speech_prob": 1.733909023304836e-11}, {"id": 40, "seek": 27580, "start": 299.8, "end": 303.8, "text": "會不會執行一半停下來?報錯之後能不能繼續?", "tokens": [51565, 6236, 21121, 16262, 115, 8082, 2257, 30018, 36135, 27769, 30, 17803, 13133, 26362, 8225, 28590, 38459, 30, 51765], "temperature": 0, "avg_logprob": -0.09867606707067297, "compression_ratio": 1.2615384615384615, "no_speech_prob": 1.733909023304836e-11}, {"id": 41, "seek": 30380, "start": 303.8, "end": 305.8, "text": "以及任務中斷之後能不能恢復?", "tokens": [50365, 40282, 26443, 39517, 5975, 4307, 115, 26362, 8225, 28590, 11845, 95, 2172, 102, 30, 50465], "temperature": 0, "avg_logprob": -0.07978242465427944, "compression_ratio": 1.1671732522796352, "no_speech_prob": 2.0962528587964435e-11}, {"id": 42, "seek": 30380, "start": 305.8, "end": 309.8, "text": "這些問題恰恰就是Harness最核心的價值。", "tokens": [50465, 29869, 17197, 11845, 108, 11845, 108, 5620, 39, 16937, 8661, 8987, 116, 7945, 1546, 45441, 40242, 1543, 50665], "temperature": 0, "avg_logprob": -0.07978242465427944, "compression_ratio": 1.1671732522796352, "no_speech_prob": 2.0962528587964435e-11}, {"id": 43, "seek": 30380, "start": 309.8, "end": 315.8, "text": "Codex Harness本身也已經把很多過去開發Agent最麻煩的事情提前處理好了。", "tokens": [50665, 34, 1429, 87, 389, 16937, 8802, 19847, 6404, 25338, 16075, 20778, 36910, 8949, 14637, 32, 6930, 8661, 38999, 17867, 102, 1546, 24675, 20949, 8945, 39289, 13876, 12621, 1543, 50965], "temperature": 0, "avg_logprob": -0.07978242465427944, "compression_ratio": 1.1671732522796352, "no_speech_prob": 2.0962528587964435e-11}, {"id": 44, "seek": 30380, "start": 315.8, "end": 319.8, "text": "比如Sandbox可以把Agent的命令執行,限制在安全環境裡。", "tokens": [50965, 36757, 50, 474, 4995, 6723, 16075, 32, 6930, 1546, 25236, 49061, 16262, 115, 8082, 11, 43446, 25491, 3581, 45466, 40063, 108, 47885, 11066, 1543, 51165], "temperature": 0, "avg_logprob": -0.07978242465427944, "compression_ratio": 1.1671732522796352, "no_speech_prob": 2.0962528587964435e-11}, {"id": 45, "seek": 30380, "start": 319.8, "end": 323.8, "text": "Approval可以讓敏感操作先經過用戶確認。", "tokens": [51165, 9132, 340, 3337, 6723, 21195, 7017, 237, 9709, 13167, 235, 11914, 10108, 15145, 8816, 9254, 1486, 114, 24293, 22041, 1543, 51365], "temperature": 0, "avg_logprob": -0.07978242465427944, "compression_ratio": 1.1671732522796352, "no_speech_prob": 2.0962528587964435e-11}, {"id": 46, "seek": 30380, "start": 323.8, "end": 327.8, "text": "Thread可以把一個完整任務和狀態保存下來,下次繼續執行。", "tokens": [51365, 2434, 2538, 6723, 16075, 8990, 14128, 27662, 26443, 39517, 12565, 18637, 222, 45248, 24302, 39781, 27769, 11, 35829, 38459, 16262, 115, 8082, 1543, 51565], "temperature": 0, "avg_logprob": -0.07978242465427944, "compression_ratio": 1.1671732522796352, "no_speech_prob": 2.0962528587964435e-11}, {"id": 47, "seek": 32780, "start": 327.8, "end": 333.8, "text": "同時還有MCP和Skills可以繼續給Agent增加外部工具和專業能力。", "tokens": [50365, 13089, 6611, 15569, 44, 20049, 12565, 50, 74, 2565, 6723, 38459, 17798, 32, 6930, 24228, 252, 9990, 12022, 13470, 23323, 39806, 12565, 46072, 27119, 8225, 13486, 1543, 50665], "temperature": 0, "avg_logprob": -0.07478047594612028, "compression_ratio": 1.0086580086580086, "no_speech_prob": 2.42918116177826e-11}, {"id": 48, "seek": 32780, "start": 333.8, "end": 337.8, "text": "如果這些能力全部從0自己開發,成本其實非常高。", "tokens": [50665, 13119, 29869, 8225, 13486, 38714, 21068, 15, 17645, 8949, 14637, 11, 11336, 8802, 14139, 14392, 12979, 1543, 50865], "temperature": 0, "avg_logprob": -0.07478047594612028, "compression_ratio": 1.0086580086580086, "no_speech_prob": 2.42918116177826e-11}, {"id": 49, "seek": 32780, "start": 337.8, "end": 341.8, "text": "而現在Codex Harness已經把它們變成了一套可以直接附用的基礎設施。", "tokens": [50865, 11070, 12648, 34, 1429, 87, 389, 16937, 25338, 42061, 4623, 26327, 11336, 2289, 2257, 1881, 245, 6723, 43297, 8842, 226, 9254, 1546, 26008, 14971, 236, 39035, 4307, 121, 1543, 51065], "temperature": 0, "avg_logprob": -0.07478047594612028, "compression_ratio": 1.0086580086580086, "no_speech_prob": 2.42918116177826e-11}, {"id": 50, "seek": 34180, "start": 341.8, "end": 353.8, "text": "這也是為什麼我覺得這次Codex Harness真正重要的地方,並不是OpenAI又發布了一個新的Agent SDK,而是開發Agent的方式正在發生變化。", "tokens": [50365, 2664, 22021, 18987, 17689, 2664, 9487, 34, 1429, 87, 389, 16937, 6303, 15789, 24928, 1546, 30146, 11, 35934, 7296, 45569, 48698, 17047, 14637, 34688, 2289, 8990, 12560, 1546, 32, 6930, 37135, 11, 11070, 1541, 8949, 14637, 32, 6930, 1546, 9249, 27584, 15789, 3581, 14637, 8244, 26327, 23756, 1543, 50965], "temperature": 0, "avg_logprob": -0.14566315445944528, "compression_ratio": 1.0944055944055944, "no_speech_prob": 2.3376288751930652e-11}, {"id": 51, "seek": 34180, "start": 353.8, "end": 363.8, "text": "以前我們做Agent往往是先選一個模型,再選LangChain,LangGraph或者其他框架,然後自己搭工具,狀態,全線,memory和工作流。", "tokens": [50965, 44996, 5884, 10907, 32, 6930, 29510, 29510, 1541, 10108, 21625, 8990, 41908, 39823, 11, 8623, 21625, 43, 656, 6546, 491, 11, 43, 656, 38, 2662, 31148, 9572, 5000, 21416, 228, 7360, 114, 11, 10213, 17645, 22220, 255, 23323, 39806, 11, 18637, 222, 45248, 11, 11319, 31699, 11, 17886, 827, 12565, 41315, 27854, 1543, 51465], "temperature": 0, "avg_logprob": -0.14566315445944528, "compression_ratio": 1.0944055944055944, "no_speech_prob": 2.3376288751930652e-11}, {"id": 52, "seek": 36380, "start": 363.8, "end": 370.8, "text": "以後則很可能變成先選擇一套成熟的Harness,然後直接在這套Harness上做自己的業務。", "tokens": [50365, 3588, 5661, 46225, 4563, 16657, 26327, 11336, 10108, 21625, 13167, 229, 2257, 1881, 245, 11336, 21805, 253, 1546, 39, 16937, 11, 10213, 43297, 3581, 2664, 1881, 245, 39, 16937, 5708, 10907, 17645, 1546, 27119, 39517, 1543, 50715], "temperature": 0, "avg_logprob": -0.06355574615019605, "compression_ratio": 1.1979522184300342, "no_speech_prob": 2.101388507647073e-11}, {"id": 53, "seek": 36380, "start": 370.8, "end": 378.8, "text": "模型可以換,業務也可以換,但底層那套讓Agent穩定執行任務的系統可以直接附用。", "tokens": [50715, 41908, 39823, 6723, 36338, 11, 27119, 39517, 47935, 36338, 11, 8395, 22816, 9636, 97, 4184, 1881, 245, 21195, 32, 6930, 13740, 102, 12088, 16262, 115, 8082, 26443, 39517, 1546, 25368, 33725, 6723, 43297, 8842, 226, 9254, 1543, 51115], "temperature": 0, "avg_logprob": -0.06355574615019605, "compression_ratio": 1.1979522184300342, "no_speech_prob": 2.101388507647073e-11}, {"id": 54, "seek": 36380, "start": 378.8, "end": 389.8, "text": "而從OpenAI目前公開的信息來看,Codex CLI,Codex App,IDE等不同形態,本身就是圍繞同一套Codex Harness能力展開的。", "tokens": [51115, 11070, 21068, 45569, 48698, 39004, 13545, 8949, 1546, 17665, 26460, 3763, 4200, 11, 34, 1429, 87, 12855, 40, 11, 34, 1429, 87, 3132, 11, 29283, 10187, 47123, 30900, 45248, 11, 8802, 19847, 5620, 2523, 235, 23141, 252, 13089, 2257, 1881, 245, 34, 1429, 87, 389, 16937, 8225, 13486, 43491, 8949, 1546, 1543, 51665], "temperature": 0, "avg_logprob": -0.06355574615019605, "compression_ratio": 1.1979522184300342, "no_speech_prob": 2.101388507647073e-11}, {"id": 55, "seek": 38980, "start": 389.8, "end": 398.8, "text": "現在真正發生變化的是,這套原本主要服務OpenAI自己Codex產品的能力,正在進一步向開發者開放。", "tokens": [50365, 12648, 6303, 15789, 14637, 8244, 26327, 23756, 24620, 11, 2664, 1881, 245, 19683, 8802, 13557, 4275, 27408, 39517, 45569, 48698, 17645, 34, 1429, 87, 33299, 30246, 1546, 8225, 13486, 11, 15789, 3581, 18214, 2257, 31429, 24282, 8949, 14637, 12444, 8949, 12744, 1543, 50815], "temperature": 0, "avg_logprob": -0.048496556282043454, "compression_ratio": 1.2165242165242165, "no_speech_prob": 1.7159506454644813e-11}, {"id": 56, "seek": 38980, "start": 398.8, "end": 408.8, "text": "所以Codex的定位也開始不再只是一個AI變成工具,它正在逐漸變成一套可以拿來開發其他智能體的通用Agent基座。", "tokens": [50815, 7239, 34, 1429, 87, 1546, 12088, 11160, 6404, 21017, 1960, 8623, 36859, 8990, 48698, 26327, 11336, 23323, 39806, 11, 11284, 15789, 3581, 2215, 238, 14065, 116, 26327, 11336, 2257, 1881, 245, 6723, 24351, 3763, 8949, 14637, 9572, 5000, 5094, 118, 8225, 23987, 1546, 19550, 9254, 32, 6930, 26008, 6346, 100, 1543, 51315], "temperature": 0, "avg_logprob": -0.048496556282043454, "compression_ratio": 1.2165242165242165, "no_speech_prob": 1.7159506454644813e-11}, {"id": 57, "seek": 38980, "start": 408.8, "end": 417.8, "text": "以後你想開發AI數據分析系統,研發Agent,辦公Agent,甚至企業內部數字員工,並不一定還要從最底層重新打一整套Agent Framework,", "tokens": [51315, 3588, 5661, 38386, 8949, 14637, 48698, 30622, 36841, 6627, 7360, 238, 25368, 33725, 11, 23230, 242, 14637, 32, 6930, 11, 23626, 13545, 32, 6930, 11, 17556, 20844, 7384, 223, 27119, 28472, 13470, 30622, 22381, 20992, 23323, 11, 35934, 1960, 24272, 7824, 4275, 21068, 8661, 22816, 9636, 97, 12624, 12560, 12467, 2257, 27662, 1881, 245, 32, 6930, 31628, 1902, 11, 51765], "temperature": 0, "avg_logprob": -0.048496556282043454, "compression_ratio": 1.2165242165242165, "no_speech_prob": 1.7159506454644813e-11}, {"id": 58, "seek": 41780, "start": 417.8, "end": 424.8, "text": "完全可以直接站在Codex Harness已經搭好的基礎設施上,繼續做自己的產品和業務能力。", "tokens": [50365, 37100, 6723, 43297, 34155, 3581, 34, 1429, 87, 389, 16937, 25338, 22220, 255, 20715, 26008, 14971, 236, 39035, 4307, 121, 5708, 11, 38459, 10907, 17645, 1546, 33299, 30246, 12565, 27119, 39517, 8225, 13486, 1543, 50715], "temperature": 0, "avg_logprob": -0.08500129618543259, "compression_ratio": 1.0836653386454183, "no_speech_prob": 1.722092086986482e-11}, {"id": 59, "seek": 41780, "start": 424.8, "end": 435.8, "text": "本期後面的實戰,我們就會直接把Codex Harness跑起來,分別演示Codex SDK和Codex App Server的使用方式,以及如何接入工具,Skills和MCP。", "tokens": [50715, 8802, 16786, 5661, 8833, 1546, 10376, 23209, 11, 5884, 47778, 43297, 16075, 34, 1429, 87, 389, 16937, 32585, 21670, 11, 6627, 16158, 31382, 25696, 34, 1429, 87, 37135, 12565, 34, 1429, 87, 3132, 25684, 1546, 22982, 9254, 9249, 27584, 11, 40282, 43526, 14468, 14028, 23323, 39806, 11, 50, 74, 2565, 12565, 44, 20049, 1543, 51265], "temperature": 0, "avg_logprob": -0.08500129618543259, "compression_ratio": 1.0836653386454183, "no_speech_prob": 1.722092086986482e-11}, {"id": 60, "seek": 43580, "start": 435.8, "end": 441.8, "text": "最後再把它真正集成進一套完整的AI數據分析Agent。", "tokens": [50365, 20578, 8623, 42061, 6303, 15789, 26020, 11336, 18214, 2257, 1881, 245, 14128, 27662, 1546, 48698, 30622, 36841, 6627, 7360, 238, 32, 6930, 1543, 50665], "temperature": 0, "avg_logprob": -0.08264711168077257, "compression_ratio": 1.146341463414634, "no_speech_prob": 1.8982787564070236e-11}, {"id": 61, "seek": 43580, "start": 441.8, "end": 452.8, "text": "同時,我們團隊也已經整理好了Codex Harness使用與開發教程,完整配置說明,以及這套基於Codex Harness重構的AI數據分析Agent項目緣碼。", "tokens": [50665, 13089, 6611, 11, 5884, 37766, 33929, 6404, 25338, 27662, 13876, 12621, 34, 1429, 87, 389, 16937, 22982, 9254, 34127, 8949, 14637, 21936, 29649, 11, 14128, 27662, 38846, 34719, 4622, 11100, 11, 40282, 2664, 1881, 245, 26008, 19488, 34, 1429, 87, 389, 16937, 12624, 43362, 1546, 48698, 30622, 36841, 6627, 7360, 238, 32, 6930, 8313, 227, 11386, 12608, 96, 16337, 120, 1543, 51215], "temperature": 0, "avg_logprob": -0.08264711168077257, "compression_ratio": 1.146341463414634, "no_speech_prob": 1.8982787564070236e-11}, {"id": 62, "seek": 45280, "start": 452.8, "end": 457.8, "text": "想直接跟著項目復線的,可以查看制定評論領取。", "tokens": [50365, 7093, 43297, 9678, 19382, 8313, 227, 11386, 2172, 102, 31699, 1546, 11, 6723, 42623, 4200, 25491, 12088, 39115, 29340, 8313, 246, 29436, 1543, 50615], "temperature": 0, "avg_logprob": -0.15580402418624523, "compression_ratio": 1.0560344827586208, "no_speech_prob": 1.5592438798450736e-11}, {"id": 63, "seek": 45280, "start": 457.8, "end": 468.8, "text": "如果之前大家學習Codex,重點還是怎麼把Codex用好,那麼從Codex Harness開始,一個更值得關注的問題已經變成了,怎麼拿Codex做出屬於自己的Agent。", "tokens": [50615, 13119, 32442, 6868, 21372, 34025, 34, 1429, 87, 11, 12624, 8216, 25583, 11614, 16075, 34, 1429, 87, 9254, 2131, 11, 20504, 21068, 34, 1429, 87, 389, 16937, 21017, 11, 8990, 19002, 40242, 5916, 14899, 26432, 1546, 17197, 25338, 26327, 11336, 2289, 11, 11614, 24351, 34, 1429, 87, 10907, 7781, 9636, 105, 19488, 17645, 1546, 32, 6930, 1543, 51165], "temperature": 0, "avg_logprob": -0.15580402418624523, "compression_ratio": 1.0560344827586208, "no_speech_prob": 1.5592438798450736e-11}, {"id": 64, "seek": 46880, "start": 468.8, "end": 470.8, "text": "Codex Harness", "tokens": [50365, 34, 1429, 87, 389, 16937, 50465], "temperature": 0, "avg_logprob": -0.7596364617347717, "compression_ratio": 0.6190476190476191, "no_speech_prob": 1.8219245556672092e-10}], "language": "zh"}