References: Agents that self-check: metacognition

Source material

Source curriculum (structural mirror, cited as further study):
• Microsoft, "Metacognition in AI Agents" (AI Agents for Beginners, Lesson 09)
  Author: Microsoft Cloud Advocates
  Lesson page: https://github.com/microsoft/ai-agents-for-beginners/tree/main/09-metacognition
  License: MIT
Clawdemy's lessons are original prose that follows the pedagogical arc of this
source. We do not reproduce or transcribe it; we cite it as the recommended
companion. All rights to the original materials remain with their authors.

Note: Microsoft Lesson 09 describes the benefits of agent metacognition without
the limits of self-reflection. This lesson supplies that counterweight
deliberately (a second look is not a guarantee), for an honest treatment.

Read this next

• Metacognition in AI Agents (Microsoft) by Microsoft Cloud Advocates. The practitioner version of this lesson, with worked examples of an agent reasoning about its own reasoning and adjusting its strategy, plus runnable samples. MIT-licensed. Strong on what reflection buys; pair it with this lesson’s “honest limit” section for the full picture.

Going deeper on self-reflection

A short, durable list. Each is a primary source for the reflection techniques behind this lesson.

• Self-Refine: Iterative Refinement with Self-Feedback (Madaan et al., 2023). The paper on a model improving its own output through rounds of self-critique. The clearest study of “generate, then review, then revise.”
• Reflexion: Language Agents with Verbal Reinforcement Learning (Shinn et al., 2023). An agent that reflects on past failures in language and uses those reflections to do better next time. The research form of the self-correction thread this lesson names.

Adjacent topics

Where this leads inside this track.

• Building trustworthy agents. The next lesson, and the start of Phase 3. The question shifts from making an agent capable to making one you can trust and ship: how it fails and what guardrails contain those failures.
• Many agents working together: multi-agent systems. The previous lesson. Reflection is the cheaper alternative to adding a reviewer agent; the two lessons are the two ways to buy reliability, one by adding an agent, one by adding a step.
• How tool use turns a model into an agent. Earlier in the track. A reflection step is strongest when it can check against a real tool result, so self-review and tool use combine.