Each Claude Code subagent carries ~436k tokens of fixed overhead before doing any work

A developer team running a weekly AI-powered product review pipeline discovered that spawning each Claude Code subagent incurs roughly 436,000 tokens of fixed overhead, regardless of the actual task size. This overhead stems from the initial context load at spin-up and a cache write on the final turn — not from the content being processed. In a controlled test, splitting a review task across three subagents cost nearly three times more than using a single agent, while the embedded product content itself accounted for only about 46,000 tokens per agent — a 9.5-to-1 ratio of fixed cost to payload. The finding challenges a common assumption that agent costs scale with content size, showing instead that the number of agents is the primary cost driver. The team recommends merging agents with overlapping roles and reserving separate subagents only for genuinely independent tasks that cannot be handled within a single context.