When you describe subagents, are those single-tool agents, or are they multi-tool agents with their own ability to reflect and iterate? (i.e. how many actual LLM calls does a subagent make?)
Now these could be anything really: API calls, pure computation, or even LLM calls.