Four ways to deploy an agent one axis: the answer window
the same agent, four answer windows — schematic, not measured
Nothing here is a preference. The window between the work arriving and the answer being due is set by the product, and it decides the rest: whether a queue can absorb a spike, whether replicas need warm models, whether there is a network at all. Pick the shape that fits the window, then live with the failure that comes attached to it.