EP449 – Claude Code 的 subagent 什麼時候該開？開之前先問這五個問題

Cloudcode's official documentation highlights the role of subagents in managing search results within main conversations. A recent article from Anthropic's Applied AI team compares various agents and suggests that a single agent with skills often outperforms multiple subagents in terms of quality, cost, and speed. The discussion emphasizes the need for careful consideration before deploying a subagent, as it introduces a new context that may not retain previous conversation details.

The episode outlines five critical questions to determine when to use a subagent. These include assessing whether the main conversation requires detailed processes or just conclusions, understanding the relationship between tasks, and evaluating the necessity of permissions. The conversation also addresses the importance of clear prompts when assigning tasks to subagents to avoid redundant work and ensure efficient output.

Additionally, the episode discusses the concept of "forking," which allows a subagent to inherit the entire context of the main conversation, making it a more efficient option for related tasks. The hosts provide practical examples of how to effectively communicate tasks to subagents and evaluate their performance over time, suggesting a systematic approach to determine their ongoing value in workflows.

Claude Code 的 subagent 什麼時候該開？官方文件說，旁邊的工作會把一大堆搜尋結果、log 灌進主對話的時候，就交給 subagent；可是 Anthropic 在 2026/10/22 的文章裡說，他們比較過的好幾家企業 agent，一個 agent 加上 skill，常常比拆給一堆 subagent 品質更好、更便宜、更快。Claude Code 自己的最佳實務文件，也把「用 subagent 保護 context」列成了舊建議。

這一集給已經在用 Claude Code 的人一套判斷方法：subagent 換到什麼、付出什麼；開一個 subagent 之前要問自己的五個問題（每一題附官方依據）；fork 是什麼、怎麼用 /subtask 開；怎麼下 prompt 把任務交代清楚（四個 Claude Code 日常情境的壞交代 → 好交代）；還有怎麼用同一批題目測試，確認一個 subagent 值得留下來。

📌 這集最重要的六個點

只要結論、不要過程（跑測試看 log、在 codebase 裡找東西）→ 開一個 subagent，要它回一份短摘要 主對話已經知道的事對這件工作有幫助（規劃完接著實作）→ 留在主對話裡做；真的要分出去，就用 /subtask 開一個 fork 主對話知道的事反而會帶偏（審自己剛寫的程式）→ 開一個沒看過對話的 subagent，只給它 spec 跟 diff 互不相干、只讀的工作 → 可以同時交給好幾個 subagent；要動手寫的工作 → 留在同一條線上 交代一個 subagent 的時候，講清楚目標、範圍、回覆格式跟邊界，它才不會做白工 只是擔心對話太長 → 先用同一批題目測試，答對的題數真的變少再開；已經在用的 subagent → 一次拿掉一個，用同一批題目比比看

⚠️ fork 的指令：agent view 開著的時候用 /subtask；agent view 關掉的時候 /subtask 不能用，要改打 /fork（agent view 開著時 /fork 是把整個 session 複製成新的背景 session）。

片中「主對話知道的事是幫忙、傷害還是無關」這個問法，是開發者分享的個人經驗。片中「三個小建議」是我自己的做法，不是官方結論。

官方文件與一手來源： ・Anthropic｜Building agents: When to fold subagents into your main agent loop：https://claude.com/blog/building-agents-when-to-fold-subagents-into-your-main-agent-loop ・anthropics/single-agent-loop-kit：https://github.com/anthropics/single-agent-loop-kit ・Claude Code 文件｜Subagents（含 fork、/subtask）：https://code.claude.com/docs/en/sub-agents ・Claude Code 文件｜Best practices：https://code.claude.com/docs/en/best-practices ・Anthropic｜How we built our multi-agent research system：https://www.anthropic.com/engineering/multi-agent-research-system ・Anthropic｜Effective context engineering for AI agents：https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents ・Cognition｜Multi-Agents: What's Actually Working：https://cognition.com/blog/multi-agents-working

參考影片： ・Ray Amjad｜Anthropic's New Subagent Rules Will Change How You Build：https://www.youtube.com/watch?v=ybx0mmd22pQ ・Rob Shocks - Switch Dimension｜Anthropic Says Delete Your Subagents：https://www.youtube.com/watch?v=KwmrSON2n-g

EasyType｜https://easytype.dev

你現在用了幾個 subagent？哪一個真的有用、哪一個可以拿掉，改成主對話裡的 skill？留言告訴我。

歡迎請我喝杯咖啡，幫助我繼續把節目做得更好唷～！ 👉 https://buymeacoffee.com/ailanrenbao -- Hosting provided by SoundOn