Why plan at all
A model that acts greedily one step at a time often loops or wanders. Planning asks the model to lay out a sequence of steps first, then execute them. This separates thinking from doing.
Common planning styles
- Decomposition split a hard goal into smaller subgoals
- Chain of thought reason step by step in natural language before answering
- Plan and execute write a full plan once, then run each step
- Replanning revise the plan when an observation contradicts it
Plan first, then act
A plan and execute agent produces a list of tasks, dispatches them, and checks whether the goal is met. If a step fails, it can replan rather than abandon the goal.
Trade offs
Upfront planning saves wasted tool calls and makes behavior easier to audit. But rigid plans break in dynamic environments, so the best agents interleave planning with execution and replan when reality diverges from expectation.
Key idea
Planning turns a fuzzy goal into an ordered, auditable sequence of steps, and replanning keeps that sequence honest when the world changes.