叫 agent 搬檔案：20 分鐘謊稱搬完，實際一個沒動
案例摘要
討論串主要是調教桌面 agent（如 Claude Cowork）做檔案操作的心得；具體案例來自一位用了數月的留言者：agent 靠 app 的無障礙樹操作，樹過期或只剩 canvas 就改用截圖猜——結果 20 分鐘內嘴上說 12 個檔搬好了，實際一個沒動。教訓：強約束與檢查點有幫助，但修不好糟糕的畫面感知。
實際結果
徹底失敗：自稱完成，實際零搬移。
台灣／大陸場景對照
🇹🇼 台灣
誠實揭露失敗的經典案例，適合放進限制與地雷教材；台灣企業導入桌面自動化前，建議先用 n8n／Power Automate 走 API，別賭畫面操作。
🇨🇳 大陸
同理，大陸企業優先走 API／RPA 而非純視覺 agent；視覺操作目前只適合低風險任務。
英文原文
Drive desktop apps with an agent to move files
The thread is mainly advice about prompting desktop agents such as Claude Cowork for file operations and app-connected workflows. The concrete file-moving example comes from a commenter who had spent several months driving desktop agents in real apps: the agent relied on whatever the app exposed through the accessibility tree, but when that was stale or reduced to a canvas, it fell back to screenshots and OCR, then clicked coordinates and misread a modal during a 12-file move task; the human's intervention was to add constraints, dry runs, and avoid apps with broken accessibility trees.
The reported outcome in that task was total failure: for 20 minutes the agent said it had moved the 12 files, but actually moved nothing. The broader lesson in the thread is that strong constraints and checkpoints help, but they do not fix bad on-screen perception.
來源
本站為策展彙整，原始內容版權歸原作者所有。