Creating subagents (& hooks) to work alongside the commands to improve [time, token, and $ cost] efficiency

[partounian]

If we use subagents & hooks, we could:

• specify the use of models like haiku to complete easier tasks much more quickly, and utilizing another context window so our main ("orchestrator") context window doesn't get bloated
• use hooks to steer the workflow/conversation with less tokens (example https://github.com/GWUDCAP/cc-sessions)
• use specific trigger phrases to force a level of thinking ("think" < "think hard" < "think harder" < "ultrathink")
  • with control over non-thinking vs thinking, we will have faster responses, better results, and less context usage
  • thinking modes also utilize prompt caching to a less efficient degree than non-thinking https://docs.claude.com/en/docs/build-with-claude/prompt-caching#caching-with-thinking-blocks

[TomLucidor]

What about the other IDEs, are there something similar to "subagents"?

[partounian]

tl;dr Yes, each to a different extent of customization

For the most part, they do have Custom Modes / Agents, but the extent & way each is implemented can vary significantly between IDEs & tools. Specifically many (or all?) have Modes/Agents, but not subagents that can automatically hand off tasks in an orchestrated fashion, or sometimes can't specify the model to use per agent. There's probably a way to get this functionality in any IDE / tool nowadays but certain IDEs might make the implementation more pain than it's worth.

Roo Code (VS Extension) has probably the closest implementation to the Claude Cade implementation in terms of functionality. (to my knowledge Roo Code pioneered that idea when they initially called it "Boomerang Mode")

[TomLucidor]

So are there approximate mapping between kilocode/roocode/cline and claude code? And maybe something similar to the CLI tools of Gemini/Qwen/GLM/DeepSeek? Also a good table on what is missing in each type of "subagent" tooling

[partounian]

Correct