Planning
Planning in AI is the problem of choosing a sequence of actions that transforms a starting situation into a desired goal, reasoning ahead about what each action requires and changes instead of reacting one step at a time. It covers both the formal field of automated planning, with precise state models and search algorithms, and the newer practice of getting language-model agents to break a task into steps and carry them out.
Classical planning formalized the problem with explicit states, actions and goals. The STRIPS planner, introduced by Richard Fikes and Nils Nilsson at SRI in 1971 for the Shakey robot project, described each action by its preconditions, the facts that must hold before it can be applied, and its effects, a list of facts added and deleted. A plan is a sequence of actions that, applied in order from the initial state, reaches a state satisfying the goal. Planners search this space with techniques such as forward state-space search guided by heuristics, plan-graph methods like Graphplan from 1995, and SAT-based encodings. The Planning Domain Definition Language, PDDL, created by Drew McDermott and colleagues for the 1998 international planning competition, standardized the notation so that planners could be compared on shared problems, and it remains the common input format of the field.
The strength of classical planning is guarantees, and its weakness is the modeling burden. Given a correct domain description, a planner can return a plan that provably achieves the goal, and optimal planners can prove that no shorter plan exists. But someone has to write the domain model in formal terms, the search can explode combinatorially, and real environments are partial, noisy and change while the plan runs. Extensions handle time, resources, uncertainty and partial observability, and hierarchical task network planning encodes how complex tasks decompose, and systems based on these ideas have been used in domains such as spacecraft operations and logistics.
LLM agent planning swaps formal models for the flexibility of natural language, and for the loss of guarantees. Instead of a domain file, an agent receives a goal in prose and generates steps. Yao et al. in 2022 introduced ReAct (published at ICLR 2023), which interleaves reasoning traces with actions and observations, so that the model thinks, calls a tool, reads the result and revises its thought; this made planning adaptive to feedback from the environment. Related methods include plan-then-execute prompting, Tree of Thoughts from Yao et al. in 2023, which explores and evaluates alternative reasoning branches, and hybrid approaches in which a model translates a task into PDDL and hands it to a classical solver.
Whether language models truly plan remains contested. Valmeekam and colleagues, in their 2022 and 2023 PlanBench work, evaluated large models on classical planning domains such as Blocksworld and reported poor performance on problems that symbolic planners solve trivially, especially when names were obfuscated, which suggests that part of what looks like planning is retrieval of familiar patterns. Reasoning-focused models have improved on such tests, but plans produced by a model should still be verified, for example by an external validator or by simulation, before they are trusted.
Pitfalls include compounding errors, over-long plans and false confidence. A fluent step list can contain an impossible step, errors in early steps propagate, and agents can loop or drift from the goal over many turns. Practical systems keep plans short, check preconditions with real tools, replan after failures, and limit steps and permissions.
Read planning as reasoning about consequences before acting: classical planners give provable plans over a hand-built model, LLM agents give flexible but unverified plans, and the sound engineering move is to pair the flexibility with checking.