Measuring the fixed overhead instead of guessing is the right move: 436K tokens of baseline changes the whole cost calculation for subagents. I've found the overhead only pays off when the subagent's task is genuinely parallel or isolatable. Did you find a task size where the overhead stops mattering?