Post

This matches what I measured. Putting one fat tool payload behind a throwaway extraction subagent cut my executor's billable tokens about 37% (before/after, n=1). The saving wasn't parallelism. It was that the orchestrator only ever saw the conclusion, never the payload.