The phrase is used for two quite different architectures. In the pipeline style, the workflow is code: a graph where each node runs an agent and passes output to the next. In the conversation-first style, the workflow is structure the agent works inside: the current step surfaces in the agent's context with its instructions, the agent does the work in its own environment, and an engine decides whether it may advance.
The practical difference shows up at the checkpoints. In a pipeline, control flow is whatever the code does. In an engine-evaluated workflow, a gate is data the engine checks on every advance: a run that has not satisfied it does not move, regardless of what the model says.
It also shows up in change management: editing one step of a stored workflow changes every future run, where editing a prompt changes only the copies you found.
Agentic workflows also differ from classic automation, which moves data between services along fixed paths. In an agentic workflow, at least some steps are judgment work done by an AI agent: read the code, draft the reply, decide whether the fix is ready. The structure around those steps is what makes that judgment repeatable from one run to the next.