The post demonstrates that in multi-agent fanout pipelines, context assembly before the LLM call — not the LLM itself — can become the dominant latency and cost driver, and that passing only compact summary structs rather than full subagent outputs resolves both problems simultaneously.