Q&A · Answers from real-world tests on X

How to Use Claude Code's Subagent (Subagent)?

Not in your language? Paste the link or the text into your AI and let it read it for you.

5 source posts · 4 creatorsLast updated Nov 3
Short answerA common approach in Claude Code is to have the main agent handle only analysis, orchestration, and acceptance, while implementation work such as reading large amounts of code, writing code, and running tests is dispatched to subagents via the Agent tool. When tasks can be split, they can run in parallel. One person had 6 subagents working at the same time and restored a complete application in 1 hour. However, Anthropic recently advised enterprises to use Subagents less, because multiple agents repeatedly handing off work loses context and slows things down. For scenarios such as parallel research, independent verification, and isolated context, they remain valuable.

How to do it

Click an avatar to see the source post
  1. 1

    In the prompt, specify that the main agent only performs analysis, orchestration, and acceptance, and that implementation work such as reading large amounts of code, writing code, and running tests must always be dispatched to subagents via the Agent tool.

  2. 2

    Some people find it most economical to set the main model to the high setting, responsible only for orchestration and acceptance, while the actual execution is handled by subagents running Opus or Sonnet.

  3. 3

    When tasks can be split into mutually independent parts, they can run in parallel. One person had 6 subagents working at the same time and restored a complete application in 1 hour.

  4. 4

    Anthropic found in enterprise projects that a single agent with Skills often works better than complex multi-agent setups. One customer switched to a single loop and became about 2x faster.

  5. 5

    Subagents still have value in scenarios like parallel research, independent verification, and context isolation. Subagents only suit very narrow tasks that can be completed independently.

  6. 6

    Before merging or removing subagents, validate with end-to-end Evals. In one case, after merging, latency dropped, but the agent missed confirming delivery time with the logistics expert, and the task failed.

What people on X saySorted by usefulness

01宝玉09-03ZHHave the Main Agent Handle Only Orchestration and Acceptance, Delegating Implementation to Subagents#1基本上你只要让 Fable 指挥其他 Agent 干活,这两个进度条就能持平。 大部分任务你选 Fable + High(或者Medium,Claude CodeOriginal post129KLikes 461
02Xudong Han10-24ZHAnthropic Recommends Using Fewer Subagents, Which Are Faster After Merging#2Anthropic 最近开始建议企业「少用点 Subagent」,甚至把一些专家 Agent 重新合并回主 Agent Loop。 他们在多个企业项目中发现,单 Agent + Skills 的效果经常比复杂的多 Agent 架构更好。Claude CodeOriginal post441KLikes 2K
03Berryxia.AI09-03ZHSubagents Easily Lose Context and Only Suit Narrow, Independent Tasks#3卧槽!!这个太难得看到了! Antropic 居然开源了他们做购物Agent,很多人第一反应是拆一堆子Agent。 Anthropic给的结论正好反过来。Claude CodeImageOriginal post101KLikes 315
04宝玉08-22ZHFable Defaults to the High Setting for Orchestration and Acceptance Only, With the Full Prompt AttachedFable 默认用 high 就够了,然后让它主要负责编排、验收,这样是最经济实惠的。 附提示词:Claude CodeImageOriginal post88KLikes 400
05WquGuru10-12ZH6 Subagents Working in Parallel Restore a Complete Application in 1 Hour用完逆向Agent REA,真的堪比原子弹爆炸,瘫坐在地上。 杀伤力多大?恐怕收藏这个项目的朋友大部分都没意识到。Claude CodeImageOriginal post85KLikes 217
No matching posts or creators found

Everything here comes from real-world tests by users on X. Rules and products change often, so always check the latest. If you know a better way, we'd love to hear it.

Related questions

How to Write CLAUDE.md?6 postsHow do beginners get started with Claude Code?5 postsHow do you save quota and tokens in Claude Code?5 postsView all 13 questions