Multi-model orchestration without the orchestra

Most “orchestration” I see is elaborate routing that exists to justify itself. The useful version is smaller and more boring.

You do not need a model router for the sake of having one. We route for exactly two reasons: cost on high-volume, low-stakes calls, and a fallback when the primary degrades. Everything else is one good model and a clear prompt.

A cascade, not a council

The pattern that earned its keep is a cascade: try the cheap model, accept its answer only if it clears a confidence and schema check, otherwise escalate to the stronger one. No voting, no debate, no five-agent committee. Each step is independently verifiable.

def answer(task):
    cheap = small.run(task)
    if cheap.confidence >= 0.9 and schema.valid(cheap):
        return cheap
    return large.run(task)  # escalate, rarely