If you don't mention it directly, it's 50/50 whether any given request will invoke a subagent.
The same with tools, skills, etc. No matter how smart these LLMs appear, they rarely do thinking as you expect.
So basically: learn how the harnesses operate, and know the names of the tools they have.