Reason and Plan

Reasoning and planning is the stage where the agent decides what to do with what it perceived: break the goal into steps, pick the next action, and choose which tool — if any — to use. This is thinking before doing. It can be as light as an implicit choice between answering and calling a tool, or as explicit as the model writing out a numbered plan, and modern reasoning models with extended thinking budgets let you dial up deliberation for hard problems.

This stage is where agent quality is won or lost. An agent that acts without planning thrashes — calling the wrong tools, repeating work, wandering away from the goal — and every wasted step costs latency and tokens. Explicit reasoning also buys you observability: when the agent states why it’s about to search the order database, you can debug a bad run by reading its rationale instead of guessing. Research like ReAct showed that interleaving reasoning traces with actions substantially improves multi-step reliability.

In practice, you elicit planning through prompting and structure. You’ll instruct the agent to “think step by step before selecting a tool,” request a plan as structured output before execution begins, or use models’ native chain-of-thought capabilities. For bigger tasks you’ll separate planning from doing entirely — the planner-executor architecture later in this module. Keep plans revisable: the reason-act cycle exists precisely because tool results often invalidate step three of a plan written at step one.