Скачать или смотреть Subagents Claude Code

We're starting to build sub-agents using Cloud Code, and this is new for me, so I'm curious to see how it goes and how useful these sub-agents really are. First, we're going to read the instructions, the manual is important, then we'll install some sub-agents in our terminal while Cloud Code runs up top. The tutorial walks us through building sub-agents, and from what I understand, Cloud Code sub-agents are basically specialized AI helpers for specific tasks with their own context windows, system prompts, and tools.

If an AI's early context is sharper, then separate agents each starting fresh should give us better output, and I like that you can fine-tune agents for things like code review, code simplification, or UX review. I set up a code simplifier agent to reduce both complexity and volume, aiming to make the code easier to read and maybe save on token costs. The setup was simple: describe what you want, let Cloud Code do its thing, pick tools, and save.

After making the agent, I ran it on a basic app and told the agent to simplify the code, and it actually worked, the file went from 298 to 196 lines, though to be fair, some changes were just merging lines. Still, having more sub-agents, like a code reviewer or a UX optimizer, could be useful, and you can set different permissions so, for example, a code reviewer cannot edit code, only comment. Right now, I see sub-agents as a way to boost productivity, reduce errors, and even make engineers happier if the tools make their lives easier.

There are other approaches, like using Genkit or LangChain for multi-agent workflows, but so far, Cloud Code makes it pretty easy to whip up a custom helper. My team and I use lots of AI every month, and while cost is a concern, the productivity gains are worth it. As more AI agents specialize, things should get cheaper and better, and even though agents are not perfect now, I'm excited about where this is heading.

If you like learning by doing, like me, trying out these tools is actually a lot of fun, and even as a non-coder, you can get value out of AI agents, just focus on giving good instructions and auditing the results.