模型更新很快，但真正要改的是你的流程
這幾天又看到一波新模型、排行榜和示範影片。畫面裡的 Agent 會自己點介面、自己修錯、自己把流程做完，看起來很厲害。
但我覺得真正值得問的，不是「它有沒有比上一個強」，而是：我手上的流程，有沒有因為它而變得更可靠？
New models, rankings, and demos arrive quickly. The better question is not “is it stronger?” but “does it make my workflow more reliable?”
我不想把模型當成一個更會聊天的工具，也不想把它當成不用看管的同事。比較接近的定位是：它可以幫我把一段工作跑得更遠，但我仍然要先畫好跑道。
I do not see a model as just a better chatbot, or as an employee I can ignore. It can take work further, but I still need to define the runway.
第一件事：不要只看單價，要看一件事怎麼完成
以前我很容易先看 token 價格，覺得便宜的模型就是比較省。但如果它在同一個任務裡反覆繞路、改來改去、最後還要我花很多時間收尾，單價再低也不是真的省。
我現在比較想記錄四個數字：
- 從開始到交付，要走幾個步驟？
- 花了多少模型成本，也花了多少人的時間？
- 交付結果有多少地方需要回頭修？
- 下次遇到同類任務，我能不能重複使用這個做法？
A cheap token is not always a cheap task. I look at steps, total cost, rework, and whether the workflow can be reused.
如果一個新模型能少掉很多來回，或把「做完、執行、看結果、再修」串成更短的閉環，它就值得我試。反過來說，如果只是摘要、翻譯或一問一答就完成的事情，不一定需要換到最強、最貴的模型。
A stronger model earns its place when it shortens the full loop: build, run, inspect, and fix. For simple tasks, the most capable model may be unnecessary.
第二件事：能力變強，權限不能跟著失控
當 Agent 可以操作瀏覽器、終端機或工作資料夾，真正要先想的不是它能做多少，而是它不該碰什麼。
- 可以自己做：整理檔案、建立草稿、跑測試、彙整結果。
- 可以做，但要在隔離環境：修改程式、建立新檔案、測試外部整合。
- 一定要人確認：發送訊息、公開發布、碰正式資料、使用金鑰或改動不可回復的設定。
As agents gain access to browsers, terminals, and files, the first question is not what they can do. It is what they must never touch without a person.
這不是不信任 AI，而是把信任放在正確的位置。能力愈強的工具，愈值得有清楚的邊界、最小權限和可以回頭檢查的紀錄。
This is not distrust. It is placing trust in the right place: clear boundaries, least privilege, and an audit trail.
第三件事：驗收比提示詞更重要
以前我們花很多時間研究怎麼把 prompt 寫得更長、更精準。現在 Agent 能自己把工作往下跑，真正重要的問題慢慢變成：什麼叫做完成？
我會先寫下驗收條件，再讓 Agent 開始做。例如：
- 輸出是不是符合既有格式？
- 有沒有改到不該改的檔案？
- 測試有沒有真的跑過，而且結果可重現？
- 如果它判斷錯了，我能不能容易復原？
As agents do more of the execution, “what counts as done?” matters more than a longer prompt.
我不需要逐行盯著它做，但也不會只看它說「完成」。我要看可以驗證的結果。
這也是人最後還不能省掉的價值：不是把每個步驟做完，而是判斷這個結果能不能真的用。
I do not need to watch every step. But I do need evidence. Human value is increasingly in judging whether the result is truly ready to use.
今晚就能做的四件小事
- 挑一個你每週都會做的任務，記錄它原本的時間、步驟和成本。
- 選一個低風險版本，讓 Agent 在獨立資料夾或測試環境完成一次。
- 在開始前寫下三條「它不能做」的規則。
- 在結束前寫下三條「我怎麼知道它做對」的驗收條件。
Pick one recurring task. Test it in a safe space. Set three limits. Define three checks.
不是每次模型更新都要立刻換工具。真正可以累積下來的，是你慢慢建立的流程、邊界和判斷標準。下個月模型名字又會換，但這些東西還會留在你手上。
You do not need to chase every release. Workflows, boundaries, and judgment are the things that compound.
Models Change Quickly. Workflows Should Last.
New models, rankings, and demos arrive quickly. The important question is not whether a model is stronger than the last one. It is whether the way we work becomes more reliable because of it.
I do not treat a model as a chatbot with better answers, or as an employee that needs no supervision. It can take a task further, but I still need to define the runway.
Measure the Whole Task
Token price is only one number. A model that takes fewer steps, needs less rework, and produces a reusable workflow may cost less for the actual job, even when its unit price is higher.
Capability Needs a Boundary
As agents gain browser, terminal, and file access, separate work they may do independently from work that needs isolation and actions that must return to a person. Stronger capability deserves clearer boundaries, least privilege, and a record that can be reviewed later.
Verification Matters More Than a Longer Prompt
When an agent can execute and correct more of the work, define done before it begins: expected format, allowed files, reproducible tests, and a recovery path. Do not only accept “completed”; ask for evidence.
Four Small Things to Do Tonight
Measure one recurring task, try a low-risk version in a safe space, write three limits, and define three checks. You do not need to chase every release. Workflows, boundaries, and judgment are the things that compound.