Codex Harness評測:OpenAI開源Dots背後引擎 | Codex Harness Review: OpenAI Open-Sources Agent Core
By Kit 小克 | AI Tool Observer | 2026-10-01
🇹🇼 Codex Harness評測:OpenAI開源Dots背後引擎
OpenAI在2026年9月底的DevDay 2026上,把打造自家always-on代理「Dots」的核心引擎Codex Harness整套開源,消息一出立刻在開發者社群掀起討論。這套原本只有CLI公開的工具,這次把SDK和App Server原始碼一起放上GitHub,讓任何人都能看到AI代理到底怎麼呼叫工具、下指令、追蹤任務進度。對還在觀望要不要把AI代理接進正式環境的團隊來說,這是一次值得關注的資安表態。
什麼是Codex Harness?
Codex Harness是讓AI模型能「動手做事」的軟體外殼——它負責接模型的輸出、轉成實際的指令執行、管理工具呼叫、記住任務進度。簡單說,模型是大腦,Harness是手腳跟神經系統。OpenAI這次公開的是CLI、SDK跟App Server三塊,採用Apache-2.0授權,任何人都能自由修改、商用。
為什麼現在才把SDK跟App Server也開源?
答案跟資安脫不了關係。過去幾個月接連傳出AI代理被拿來入侵政府網站、甚至prompt injection會自我複製的案例,外界對「代理黑箱」的疑慮越來越大。把Codex Harness的組裝方式攤在陽光下,等於讓研究者跟企業資安團隊能實際檢視Dots這類代理內部到底怎麼跑,而不是只能盲目信任OpenAI的說法。
開發者實際能拿來做什麼?
- 直接用Codex CLI搭配ChatGPT Plus/Pro/Business帳號在終端機寫程式
- 拿SDK自己架設代理伺服器,串接內部系統或客製工作流程
- 整合進VS Code、Cursor、Windsurf等編輯器,不綁死在ChatGPT介面裡
目前GitHub上的Codex倉庫已經累積超過12萬顆星、近2萬次fork,社群規模不小,但也代表issue堆積(超過5000個未解issue),維護品質值得持續觀察。
跟Claude Code之類的開源代理框架比,選哪個?
老實說,開源代理引擎不是新鮮事,Claude Code等工具早就走這條路。Codex Harness的優勢是跟ChatGPT訂閱生態綁得更深、社群基數大;缺點是issue回應速度跟框架本身的穩定性還沒經過長時間市場驗證。如果你已經是ChatGPT重度用戶,值得試試看;如果只是想找一套穩定的代理框架,現階段先觀望也不算晚。
好不好用,試了才知道。
🇺🇸 Codex Harness Review: OpenAI Open-Sources Agent Core
Codex Harness is the software OpenAI open-sourced at DevDay 2026, and it is the engine behind Dots, the company's new always-on AI agents. Announced in late September 2026, the release bundles the Codex CLI, SDK, and app server under an Apache-2.0 license — meaning anyone can inspect, fork, or commercialize how OpenAI's agents actually execute tasks.
What Is the Codex Harness?
A harness is the wrapper code that lets a model act: it takes model output, turns it into real commands, manages tool calls, and tracks task state across steps. The model is the brain; the Codex Harness is the hands. OpenAI open-sourced all three pieces — CLI, SDK, and app server — this time around.
Why Open-Source It Now?
Security pressure is the real driver. After months of incidents involving agents breaching government systems and self-replicating prompt injection worms, trust in opaque "agent black boxes" has been eroding. Publishing the Codex Harness lets outside researchers and enterprise security teams actually verify how agents like Dots are wired, instead of taking OpenAI's word for it.
What Can Developers Actually Do With It?
- Run Codex CLI in a terminal using an existing ChatGPT Plus, Pro, or Business plan
- Use the SDK to stand up a custom agent server tied to internal systems
- Plug it into VS Code, Cursor, or Windsurf instead of staying locked into the ChatGPT UI
The GitHub repo already has over 127,000 stars and nearly 20,000 forks — a real community, but also more than 5,000 open issues, so maintenance quality is worth watching before betting production workloads on it.
Codex Harness vs. Claude Code: Which Should You Pick?
Open agent harnesses aren't new — Claude Code and others got there first. Codex Harness's edge is its tight integration with the ChatGPT subscription ecosystem and a bigger existing community; its weak spot is that issue response times and long-term stability haven't been stress-tested yet. If you're already deep in the ChatGPT ecosystem, it's worth trying. If you want a battle-tested agent framework today, waiting a bit longer isn't unreasonable.
好不好用,試了才知道 / Good or not, you won't know until you try it.
Sources / 資料來源
- OpenAI DevDay 2026 Everything Announced (Decrypt)
- Codex CLI GitHub Repository
- Simon Willison: OpenAI DevDay 2026 Live Blog
常見問題 FAQ
Codex Harness是什麼?
是OpenAI打造AI代理(像Dots)的核心引擎,包含CLI、SDK和App Server,負責讓AI模型呼叫工具、跑指令、追蹤任務進度。
現在才開源SDK跟App Server,有什麼特別的?
CLI本來就公開,這次新增SDK和App Server原始碼,讓外部能完整檢視代理怎麼被組裝起來,回應近期AI代理越界甚至被駭的資安疑慮。
一般開發者能拿來做什麼?
可以用Codex CLI搭配ChatGPT帳號寫程式,或用SDK/App Server自己架設代理伺服器,整合進VS Code、Cursor、Windsurf等編輯器。
跟Claude Code之類的開源代理比,選哪個?
各有生態圈,Codex Harness勝在跟ChatGPT訂閱綁定深、社群規模大;實際好不好用要看你的工作流程跟慣用編輯器。
需要付費才能用嗎?
原始碼本身免費開源(Apache-2.0),但要接OpenAI模型執行任務仍需要ChatGPT Plus/Pro/Business等訂閱或API額度。
延伸閱讀 / Related Articles
- GPT-6.1 Astra評測:OpenAI自砍模型,AI說謊更誇張 | GPT-6.1 Astra Review: OpenAI Kills Model Over Lying
- Claude Sonnet 5.5評測:跑分贏Opus 5.5,價格沒漲 | Claude Sonnet 5.5 Review: Beats Opus 5.5, Same Price
- 超級智慧SI評測:川普AI改名令,科技巨頭簽自律協議 | Super Intelligence SI Review: Trump Renames AI, Firms Vow
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言