The Parallel Processing Paradox: Why More Subagents Slowed My Pipeline

Adding more parallel subagents paradoxically increased latency—but not where expected. The bottleneck wasn't LLM inference; it was context assembly consuming over half the wall-clock time before the model was even called. By rethinking data flow with R2 storage and compressed summaries, latency improved and monthly costs fell from $207 to $38, revealing how aggregation overhead—not raw parallelism—often limits distributed architectures.