Responding to now vs. decomposing a goal in advance
A reactive agent looks only at its current state and decides its next single action, with no explicit representation of the steps beyond that. It's simple: perceive, decide one action, act, repeat. This works well for narrow, well-defined tasks where the right next step is usually obvious from the current state alone, like answering a single lookup question or responding to one well-scoped support ticket.
A planning agent instead performs task decomposition: before taking any action, it breaks a larger goal into a sequence of steps, for example, 'to book this trip I need to: find flights, compare prices, check the user's calendar, then confirm.' It commits to that sequence, or at least a rough one, before executing the first step, which is fundamentally different from deciding one action at a time.