In plain terms
A builder renovating a kitchen starts with a list, and the hammer comes later. She writes down the jobs, puts them in order because the plumbing has to go in before the tiles, and changes the order when the wall turns out to hide a rotten pipe. An agent that plans does the same: it writes down the steps to the goal, works through them and rewrites the list when reality disagrees. The plan is a working document that anyone can read.
Why it matters
Planning is what lets an agent finish a forty-step task where another loses its way after ten. For a manager it has a second use: a written plan is the cheapest place to supervise. Reading it before execution takes minutes and can prevent hours of work in the wrong direction, which is why many tools have the agent stop for approval at that point. The limits are real. A plan rests on what the agent knew at the start, it can sound convincing and still be wrong, and on small tasks planning only adds cost and delay.
Example
A coding agent is asked to move a billing service to a new database. It first writes a five-step plan: back up, inventory the 14 tables, migrate a copy, compare row counts, switch over. A lead engineer reads it in three minutes and adds a rollback test before the switch. At step four the counts differ by 212 rows. The agent adds a step to find the cause, fixes a date-format error and continues.
Most often confused with
Planning vs. Chain of Thought (CoT)
Both are a model thinking in steps, which is why they get mixed up. Chain of thought is reasoning inside one reply: the steps are thoughts, and nothing happens in the world. A plan is a list of actions to be carried out over many turns, each of which can fail and force a change. A reasoning model makes better plans; the plan still has to be executed, checked and updated.
Under the hood
Planning comes in forms from light to heavy. Implicit: in a ReAct-style loop the model decides only the next step. Explicit: plan-and-execute designs have a planner write all the steps, an executor carry them out and a replanning step update the remainder after each result. Hierarchical: a lead agent splits the goal into sub-goals and assigns them to subagents. Many harnesses give the model a to-do list tool, so that the plan lives as an explicit item in the context or in a file and survives compaction. A plan mode restricts the agent to read-only exploration until a person approves the plan. Good plans say what done looks like and include a check for each step. Pitfalls: plans built on assumed facts, such as files that do not exist; plans followed after they have gone stale; splitting a task too finely; and drift from the original goal on long runs. A common cost pattern has a stronger model plan and a cheaper one execute. Evaluate the plan and the execution separately.