Metacognition is "thinking about thinking": the agent evaluates and adjusts its own reasoning. It can reflect on past choices ("I prioritized cost; I might have missed direct flights"), adapt strategy from feedback ("When the user says ‘too crowded,’ I should change my ranking method"), and correct errors (check availability before suggesting a hotel). Techniques include planning with clear steps, corrective RAG (retrieve, then evaluate and correct), and using code or SQL as tools with schema-aware reasoning.
Metacognition: reflect → adjust flow
Metacognition: thinking about thinking
Self-reflection — Agent assesses its own choices: "I picked the cheapest flight; I might be missing direct options. Let me re-check."
Adaptation — Notices patterns: "When the user says ‘too crowded,’ I should change how I rank attractions, not just remove one."
Error correction — Learns from failure: "Last time that hotel was fully booked; I’ll check availability before suggesting."
Example: Hotel recommendation
Agent first picks the cheapest hotel. User feedback: "bad" (e.g. quality too low). The agent reflects, switches strategy to "highest_quality," and recommends again. That reflection-and-adjust loop is metacognition in action.