Agent Basics 09 Gets Agents to Check Their Own Work

Metacognition means getting an Agent to do more than give an answer. It also looks back and asks: why did I choose this, did I go wrong somewhere, and should I try a different approach next time.

Remember These Points First

• True metacognition means reasoning about your own reasoning, not just revising the final answer.
• The benefits: more transparency, better reasoning, the ability to adapt to change, and more accurate self-awareness.
• Four uses: self-reflection, adaptation, error correction, and resource management.
• An Agent is made up of persona, tools, and skills. Metacognition adds a layer of "review" on top of them.

Going Deeper

An example. A travel Agent recommends many popular attractions, and the user says "it's too crowded." The usual fix is to delete a few attractions. The metacognitive fix is: "I've been sorting by popularity, and that method itself is the problem. Next time I should consider crowd levels." It changes the selection method, not just this one result.

Common practice. After giving an answer, have it self-check first: what is the basis for this conclusion, what did I miss, is there a better option? Then decide whether to redo it. When the same mistake repeats, record the lesson so it can be applied directly next time.

Caution. Frequent review adds time and cost. Only important tasks and error-prone steps are worth thinking through twice.

Image from the course repository.

Try It Yourself

Take an answer the model just gave you and have it review the answer under three points: basis, omissions, and a better approach. See whether it can produce a better version.

Let an Agent Teach You

Paste the following into your Agent (Claude Code, ChatGPT, and Gemini all work), and it will guide you step by step like a mentor.

References

• Content adapted from Lesson 9 (09-metacognition) of microsoft/ai-agents-for-beginners, originally released under the MIT License.
• For the original and full code: 09-metacognition
• For all courses, see the Agent course roadmap