What is Agent Planning?

Without planning, an agent operating purely reactively can wander — taking a locally reasonable next step that doesn't actually serve the overall goal, or looping between two states without making progress. Planning gives the agent (and anyone reviewing its work) a structure to check against: does this sequence of steps actually lead to the stated goal, and does it visibly account for dependencies, like needing to gather information before it can act on it. Plans in agent systems are usually not fixed once and executed blindly — they're revised as new information comes in mid-task, the same way a human plan adapts when a first attempt reveals a wrong assumption. This is different from planning in classical AI, where a full plan was computed up front and executed rigidly; modern agent planning is closer to "plan, act, observe, replan," which is also why a visible plan is useful for a human reviewer to sanity-check before costly or irreversible steps are reached.