Planning is how an agent decomposes a complex goal into executable sub-tasks and determines the order of steps. It is one of the most important factors in whether an agent reliably completes complex tasks or gets lost mid-execution.
The model decides what to do at each step from the accumulated context. No upfront plan. Works for short tasks. Degrades on longer tasks where losing the overall goal is easy.
The model produces a structured list of sub-tasks before acting. Forces thinking through the entire task upfront — catching dependency issues and scope gaps. More reliable for complex tasks.
Chain-of-Thought (CoT) — The model reasons step-by-step in the Thought phase before each action. Reasoning is visible and correctable. Significantly improves reliability. The standard approach for most agents.
Tree-of-Thought (ToT) — Generate multiple candidate next steps, evaluate each, select the best. More compute-intensive but better for tasks with many possible solution paths — complex problem solving, strategic planning.
Plan revision — When a step fails or new information contradicts assumptions, the agent revises its plan rather than continuing blindly. Requires holding both "original plan" and "current state" in context.
Reflection — After completing a task or at key milestones, the agent evaluates the quality of its own outputs and identifies potential errors before proceeding. A self-critique loop that catches silent mistakes.
Ask the AI assistant about agent planning, chain-of-thought reasoning, tree-of-thought, or how to implement plan revision.