If you ask Claude to spawn an agent, it will create a Task() with its own context window. It seems like the tokens used by these subagents are not tracked, rather only the token use from the primary Claude agent? (based on looking at the token report input/output tokens).
my token report:
