What is planning in AI?
Planning in AI, formally called automated planning, is the problem of choosing a sequence of actions that takes a system from its current state to a goal state. Each action has preconditions (what must be true to take it) and effects (what changes afterward). A planner searches over possible sequences, discards those that violate constraints, and returns an ordered set of steps. Classical planners assume a fully known, predictable world. Extensions handle uncertainty, actions that may fail, time and resource limits, and goals that change during execution.
Planning differs from scheduling, which assigns already-chosen tasks to times and resources, and from optimization, which finds the best values for a set of decisions. In practice the three are combined: a planner decides what to do, a scheduler decides when, and an optimizer decides how to do it at the lowest cost.
How planning works
Traditional planners use symbolic search with heuristics to prune the enormous space of possible action sequences, and hierarchical planners break a goal into subgoals that are planned separately. In modern AI agents, a large language model often serves as the planner. Given a goal, the available tools, and the current state, it proposes the next step, executes a tool call, reads the result, and revises. This loop, described by the ReAct pattern, is less rigorous than classical planning but handles ambiguous, language-heavy tasks that formal planners cannot represent.
Planning in practice
Planning is the capability that separates an agent from a chatbot. An agent takes a goal such as “onboard this new supplier,” works out that it must validate the tax ID, create the vendor record, request banking details, and route the contract for signature in that order, and stops for approval where policy requires. A third-party logistics company uses planning at both levels. In the warehouse, a classical planner sequences the moves of autonomous carts so that each reaches its pick location without deadlock. In the office, an agent handles exceptions: when a shipment misses a connection it plans the recovery (rebooking, notifying the customer, filing a claim) and presents the plan to a coordinator for approval before any action is taken.
What to watch for
- Require the agent to show its plan before executing any step that changes a system of record.
- Set limits on plan length, retries, and spend, since an agent that keeps replanning can loop indefinitely.
- Validate language model plans against explicit rules where correctness matters, instead of trusting the model’s judgment alone.
- Test with goals that are impossible or ambiguous to confirm the planner fails safely and asks for help.