## Definition **Agent planning** is the act of producing a plan — decomposing a goal into an ordered set of subtasks — before or while executing it. Rather than letting an agent improvise one step at a time, planning makes the intended trajectory explicit so it can be inspected, validated, and corrected before any tool fires. ## Decouple planning from execution The single most useful structural decision is to separate the *planner* from the *executor*. The planner reads the goal and emits a sequence of steps; the executor carries them out. This buys you a checkpoint: the plan exists as data before the world is touched. Huyen argues in *[[AI Engineering - Chip Huyen]]* that you should **validate the plan before running it** — and that validation can be cheap heuristics (does every step reference an available tool? are there obvious ordering violations?) or an AI judge that critiques the plan as a whole. ```text goal ──▶ [planner] ──▶ plan ──▶ [validator] ──▶ [executor] │ reject └──▶ replan ``` Catching a doomed plan at the validator costs one extra call; catching it after ten executed steps costs ten irreversible actions plus the cleanup. ## Planning is search, and search is hard A plan is a path through a space of possible action sequences, so planning *is* search — and the space explodes combinatorially. This is why planning is the part of agent design that most often disappoints: models propose plausible-looking plans that are subtly infeasible, miss a dependency, or assume a tool that doesn't exist. Strong planning correlates with strong models, and even then it benefits from structure like [[Tree of Thought]] to explore alternatives, or [[Reflection]] to critique and revise a draft plan. ## The intent / scope step A frequently overlooked first move is an **intent step**: before planning *how*, decide *whether*. The agent classifies the incoming request and can mark it out-of-scope, ambiguous, or refused. This prevents the agent from confidently planning its way through a task it should never have accepted, and it gives a clean place to ask the user a clarifying question instead of guessing. ## Plan, then loop Planning does not replace the [[Agentic Loop]] — it seeds it. A common shape is: plan up front, execute steps, and after each step re-evaluate whether the remaining plan still holds (replan on surprise). The [[ReAct Pattern]] is the degenerate case where the plan is one step long and regenerated every turn; explicit planning is what you reach for when one-step-at-a-time wanders. For hard goals, planning is also the natural seam along which to split work across a [[Multi-Agent System]], with a planner-orchestrator handing subtasks to specialist executors. ## Related - [[Agentic Loop]] - [[Reflection]] - [[ReAct Pattern]] - [[Tree of Thought]] - [[Multi-Agent System]] - [[AI Engineering - Chip Huyen]]