Why split a RAG pipeline into a controller plus sub-age...

A common misconception is that agents are inherently more modular or accurate. In reality, splitting a large context into m agents reduces the prefill time by dividing the attention cost, but each agent must then decode a summary, which is a serial process. The compute win is often a wash; the true benefit of an agentic architecture is the ability to perform iterative, overlapping retrieval and re-querying, which a single long-context call cannot do.