Because criticism is only useful when it comes from outside the reasoning that produced the work. An agent carrying the generator's full trace has already absorbed its assumptions, its interpretation of the task, and the dead ends it rejected — so it tends to validate rather than challenge. Given only the artefact and the criteria, it evaluates what is actually there, which is what you wanted. Independence is the mechanism, not the model.

Because the classification call is small and the work it avoids is large. Real traffic contains many simple requests that do not need tools, a long context, or a strong model. Paying a tiny routing cost on every request to send a substantial fraction to something cheap nets out negative. It also improves the hard cases, since each path gets a focused prompt and a short tool list instead of one prompt covering everything.

They are not five perspectives, they are one perspective purchased five times. Genuine diversity comes from different tools, different data, different instructions, or an adversarial role — not from labels. Without one of those, the agents converge on similar answers, and any apparent agreement is a measurement of the same model agreeing with itself, which is worse than useless because it looks like corroboration.

Because everything downstream depends on its plan, and its errors are silent and expensive. If it splits the task badly, every worker executes correctly against the wrong brief, and the failure only surfaces at assembly — after you have paid for all of them. It is also where duplicated work originates, since overlapping subtask definitions send two workers after the same thing. Traces at the planning step are worth more than traces anywhere else.