Token accounting and dollar accounting diverge here in a way that softens the headline without touching the conclusion. The measurement that "each subagent costs roughly 436,000 tokens in fixed overhead before it does any useful work" sums usage fields where cache reads and cache writes count at face value, but cache reads bill at a tenth of base input and writes at 1.25x, and spin-up overhead is overwhelmingly cached system prompt and tool schemas. So the 9.5:1 fixed-to-payload ratio in tokens is likely closer to low single digits in dollars, and it would be worth publishing the cache-read versus cache-creation split from those same transcript fields, since they are right there. What survives untouched, and matters more: headcount is still the lever, overlapping reviewers still pay the fixed cost twice to find the same defect twice, and "spawn for independence, not for tidiness" is a one-line subagent policy worth adopting wholesale. The fabricated-quote catch by the one agent that fetched primary sources is a quietly damning result for narrow-perspective review fleets.