Planning

In artificial intelligence, automated planning is the branch that, given an initial state of the world, a goal to reach and a repertoire of available actions, searches for a sequence of actions —a plan— that turns the initial state into one where the goal holds. Knowing what we want is not enough: we have to find the concrete chain of steps that gets us there.

What sets planning apart from plain search (see that entry) is that the actions are not listed by hand but described declaratively: each one states what must be true for it to run and how it changes the world. From that description the planner derives the space of possibilities on its own. The canonical reference is still the textbook by Stuart Russell and Peter Norvig, «Artificial Intelligence: A Modern Approach».

The ingredients: states, actions and the STRIPS and PDDL languages

A planning problem is built from three pieces. The state describes how the world stands at a given moment; the goal, the condition we want to hold; and each action is defined by its preconditions —what must be true to apply it— and its effects —what becomes true or false once applied—.

This model took shape in 1971 with STRIPS, the planner built by Richard Fikes and Nils Nilsson at the Stanford Research Institute, which represented every action with an add list and a delete list. From that idea descends PDDL (Planning Domain Definition Language), the standard language that Drew McDermott and colleagues proposed in 1998 for the International Planning Competition, and which ever since lets planners be compared on shared problems.

How it is solved: heuristic search and its challenges

Once the problem is described this way, planning becomes a matter of searching within a state space: exploring which actions to apply, and in what order, until a sequence reaches the goal. The obstacle is combinatorial explosion, because the number of states grows overwhelmingly and blind search is unworkable.

What made modern planning practical was guiding that search with heuristics derived automatically from the problem itself. Blai Bonet and Héctor Geffner showed how to estimate the distance to the goal by solving a relaxed version —ignoring, for instance, the effects that undo facts already achieved—. Open challenges remain in environments with uncertainty or only partial observability, where the agent does not know the true state of the world: conformant planning and planning under uncertainty are still far costlier than the classical case.

From classical planning to LLM agents

For decades planning was symbolic and detached from learning. Around 2026 the pendulum swings back: agents built on large language models embody a form of planning when they break a goal into intermediate steps. That flexibility revives an old ambition —marrying the robustness of symbolic planners with the breadth of neural models—, for example letting the model translate the problem into PDDL and letting a classical planner guarantee a sound plan. It is still unsettled ground; for a fuller treatment we point to the entry on agents.