An agent flow is deterministic: the same input produces the same output, every time. That sets it apart from an agent's more open-ended reasoning. Every flow needs a trigger, the event that starts it, and at least one action, a task it performs. The designer also supports conditions, loops, child flows, human approval steps and AI actions that call a language model mid-flow.
A flow can run standalone, for example on a schedule, or be attached to an agent as a callable tool through the "When an agent calls the flow" trigger. That second pattern turns an agent that talks about your process into one that acts on it: the agent calls the flow, gets a live result back and uses it in its reply. Test runs in the designer do not use Copilot Credits; live runs do.