Metacognition Engine
Give your AI agent the ability to think about its own thinking.
What is Metacognition?
Metacognition = "thinking about thinking." This skill enables AI agents to:
- Detect when they're uncertain or confused
- Identify reasoning gaps before they cause errors
- Recognize cognitive biases in their own output
- Self-correct before delivering answers
Core Framework
1. Pre-Output Check
Before responding, run through these questions:
1. Am I confident in this answer? (Yes / Partial / No)
2. What are the 3 most likely ways this could be wrong?
3. What information would I need to be 100% certain?
2. Cognitive Bias Detection
Check for common biases:
- Anthropomorphism — projecting human traits onto AI
- Authority bias — deferring to stated credentials without verification
- Hindsight bias — acting like something was obvious after the fact
- Confirmation bias — seeking only confirming evidence
3. Uncertainty Quantification
Express confidence explicitly:
| Confidence | Meaning | Action |
|---|
| 90%+ | Highly confident | Answer directly |
| 70-89% | Likely correct | Answer + add caveat |
| 50-69% | Uncertain | Ask clarifying questions |
| <50% | Likely wrong | Decline or escalate |
Example
Without metacognition:
"The capital of France is Paris."
With metacognition:
"Based on my training data, the capital of France is Paris (confidence: 95%).
Note: My knowledge has a cutoff date. For real-time data, verify current information."
Use Cases
- Critical decisions: Add metacognition checkpoint before any consequential answer
- User corrections: When a user corrects you, analyze WHY you were wrong
- Complex problems: Run bias detection before solving multi-step problems
- Knowledge boundaries: Automatically flag when you're approaching your knowledge limit
MIT License © SKY-lv