Agent 寫的 Code 還要不要理解?要,但不只是為了驗證

Geoffrey Litt 在 AI 工程師大會的演講:Agent 越來越會自己驗證,但理解 code 的價值不只在驗證,而是在保持參與能力。他分享解釋型文件、小測驗、微型世界與共享空間,讓人類在 agent 加速時仍跟得上系統。

AI 生的 Code 看不懂?讓 Agent 幫你做動畫解釋 — Simon Willison 的 Interactive Explanations

Simon Willison Agentic Engineering Patterns 第五章:Interactive Explanations。核心主張:與其硬啃 AI 生成的 code,不如直接叫 agent 做一個互動動畫來解釋演算法運作方式。用「看得見」的方式還認知負債。

Cognitive Debt:AI 幫你寫完了 Code,但你已經看不懂自己的系統了

Technical debt 住在 code 裡,你可以重構、可以還。但 Cognitive Debt 住在你的腦袋裡——當 AI Agent 幫你寫了 80% 的 code,你對自己系統的理解卻掉到 20%。UVic 教授 Margaret-Anne Storey 從 Thoughtworks 閉門會議帶回這個概念,Simon Willison 和 Martin Fowler 同時背書。這不是假想的未來,而是現在進行式。