Original video in Chinese.
Key Takeaway
- Advantages of Sub-agents: create task-specialized agents, with independent context to avoid contamination and reduce cost; flexible configuration of permissions, tools, project/global scope; supports workflows and improves handling of complex coding tasks.
- Creation and use: simple steps (/agents command, description, tools, color); invoked in natural language, color-coded; examples: code review and security check agents.
- Strategic importance: Claude Code’s most critical feature, from lone operator to coordinated force; better than community projects like SuperClaude (optimization, integration); core capabilities + Sub-agents make it possible to lead a team, and this can be generalized to other fields.
The new Sub-agents feature in Claude Code is very useful, and also very important.
Simply put, you can create a bunch of Agents through this feature. Each Agent specializes in one particular task. For example, some are dedicated to code review, and some are dedicated to front-end work.
To let these Agents focus on their work, they each have their own independent context window. I think this design is fantastic, and I especially like it!
You can think of the context window as a workbench. Each Agent has its own workbench, so when they’re working, they won’t get mixed up with others, which prevents “memory contamination.” And compared with everyone sharing one huge workbench, giving each Agent a workbench that best fits its workload is the most cost-effective allocation method.
These Agents are very flexible in terms of configuration. You can set their permissions, such as which tools they’re allowed to use.
This is pretty important too. Because if there were only one Prime Agent, once there are too many tools, it may not be able to choose the most appropriate one. Sub Agents don’t have that problem; they just use the small number of assigned tools to complete the task.
In addition, you can also decide whether an Agent serves the current project or serves globally.
Finally, these Sub Agents can also form a workflow. You can orchestrate them through custom commands.
I think Sub-agents may be Claude Code’s most important feature.
With it, Claude Code has evolved from a lone operator into a coordinated force, capable of handling much more complex, large-scale tasks.
And this whole system can also be generalized to more fields, not just programming.
So this is absolutely a strategic new feature.
Hello everyone, welcome back to my channel. Humbly speaking, I’m one of the few bloggers in China who can explain both the Why and the How of AI clearly. What I offer is far more valuable than tutorials. Remember to follow me. If you want to connect with me, come join our newtype community. This community has been running for 500 days, and more than 1,500 people have paid to join.
Back to today’s topic: the Sub-agents feature.
Creating Sub-agents is very simple. Anyone can do it.
Step one, type /agents, then choose “Create.” At this point, the system will ask whether you want to create it for this project or globally. I’m choosing to create it globally.
Step two, the system asks whether you want Claude to create it for you or create it manually. I choose to let Claude help create it.
Step three, we need to provide a description of this Agent. Everyone should be very familiar with this—it’s a lot like the system prompts I often write before. I’ll keep it simple here and just ask it to help me create an Agent for code review.
Step four, decide which tools to equip the Agent with. You can choose all of them, or carefully pick specific ones.
Step five, choose a color scheme for the Agent. The reason for doing this is that in a long stream of running prompts, you can clearly see which Agent is working.
Finally, the system will summarize everything from the previous steps for us to confirm. After pressing Enter, the Agent is created.
Using the same method, I created another Agent specifically for reviewing security vulnerabilities.
Now let’s test it. In the chat box, directly tell Claude Code in natural language to use these two Agents for code review and security vulnerability detection. You can see that the Agent we just set up with a color has been activated.
After Sub-agents came out, my first reaction was that it’s very similar to a community open-source project called SuperClaude. I was originally planning to make a video introducing this project. Their core ideas can be said to be completely the same:
Using a divide-and-conquer strategy, break a big task into smaller tasks, then let multiple Agents work together to solve one large, complex problem.
The difference is that one is an official native feature, while the other is a community framework.
Because it’s official and native, there must be deep optimization on the backend, reducing unnecessary API calls and data transfer, making it faster and cheaper.
Because it’s official and native, the integration is better, and stability and reliability are higher.
So after Sub-agents came out, these community tools all have to find a way to move up another level.
Looking back, Claude Code had been laying the groundwork all along:
Claude, as the underlying model, is able to perform complex reasoning.
The long context window provides cognitive bandwidth for understanding large projects and complex requirements.
More than a dozen built-in tools and MCP external tools allow the AI to actually operate the terminal and read and write files.
And Hooks are the guarantee of discipline, ensuring the AI doesn’t go off the rails.
All of this gives Claude Code the foundational abilities of an elite programmer. Now, Sub-agents are here. They give the elite programmer the ability to lead an entire team in coordinated action. That’s why I say it’s Claude Code’s most important feature so far.
What I’ve introduced here are all basic uses. I’ll share more advanced uses in the community.
OK, that’s it for this episode. If you want to understand AI, want to become a super individual, and want to find like-minded people, come join our newtype community. See you next time!