Agents are for ambiguous work
An agent shines when the work needs judgment, when inputs vary, and when the right tool for a given step is not knowable in advance. The cost is higher and the observability is harder. The benefit is that ambiguous work that used to require a person now scales.
Automation is for well-shaped work
Process automation shines when the steps are well-shaped, the inputs are predictable, and the rules are knowable. The cost is lower, the observability is straightforward, and the runtime is cheap. The constraint is that the work has to fit the shape.
Most real systems use both
The interesting systems we ship use both. Process automation handles the well-shaped path. Agents handle the ambiguous one, with the automation feeding inputs in and consuming outputs. The architecture question is which layer holds which decision.