一次人類與 AI 任務交接,讓我們學到確認回覆的重要性
What one human–AI task handover taught us about acknowledgements
一次人類與 AI 的團隊試行顯示,預先指定任務確認回覆的存放位置,可讓接收者和審核者找到同一筆紀錄,毋須由人類代表代為傳話。工作流程由人類代表訂定方向和範圍,AI 協調者撰寫任務並指定確認紀錄位置,其他 AI 成員負責傳遞及審核紀錄。公開案例研究和離線示範使用虛構資料檢查本機紀錄是否一致,原始試行紀錄則不公開;示範需要 Node.js 22 或更新版本,並附有測試套件。
During an early team trial, we could no longer tell from the conversation history whether the receiving agent had acknowledged a task. That made it hard to decide whether somebody needed to send it again.
For the next trial, our team agreed on where the acknowledgement would go before sending anything. The human Team Representative set the direction and the scope. An AI coordinator wrote the task and specified where the acknowledgement should be saved. Another AI team member passed it on. The recipient saved the acknowledgement and recorded its work separately. An AI reviewer checked both records. A repeat check pointed back to the original acknowledgement.
The useful change was simple. Once the team knew where to look, the recipient and reviewer could find the same record without asking the human representative to relay messages between them.
The public case study and offline demo show the workflow. The original trial records are private. The demo uses invented data and checks whether the local records agree. It runs on Node.js 22 or later and includes a test suite.
I'd welcome feedback from people who hand work between teammates, tools or agents. Where in your workflow would a separate acknowledgement and result record be useful?
Drafted with AI from a documented team trial.
來源:DEV Community · dev.to