Reflection is a meta-cognitive pattern where an agent pauses to evaluate its own prior output or reasoning before proceeding. The agent may re-read its last response, check it against the original goal, identify logical errors or missing information, and decide whether to revise, retry, or continue. Reflection improves output quality in complex tasks by catching mistakes that a single-pass generation would miss, and is a key component of iterative refinement loops in agentic systems.