Recipe Overview

LLMs may make logical mistakes without self-review. A reflection agent periodically asks the model to examine its own reasoning. The problem it solves is unchecked errors in chain-of-thought. In this pattern, the agent instructs the LLM to articulate its thought process and then verify or correct it. For example, after generating a solution, the agent might ask 'Is this reasoning sound?' or 'What could go wrong with this approach?' This metacognitive step catches errors and improves answer reliability, especially for complex reasoning tasks where initial intuition might be flawed.