metacognition

To build a feedback loop for AI agent self-improvement, this skill converts human corrections and experiences into evolving perceptions, behavioral guardrails, and self-observations. It strengthens relevant entries through reinforcement and prunes outdated ones using time-based decay.

Metacognitive tracking for autonomous agents records what an agent knows about itself, its protected behaviors, and decision confidence. It maintains a self-model that uses evidence and curiosities to continuously adapt operational constraints and perceptions.

To create behavioral guardrails from agent failures, you record overrides during error-correction loops. These overrides become non-negotiable behavioral constraints that prevent the agent from repeating the same failure in future workflows.

Yes, you can use a root SKILL.md entry point with YAML frontmatter for agent metacognition. The skill compiles active metacognitive entries from memory storage into a lens for BOOT.md injection, enabling continuous self-correction.

Yes, confidence tracing works for auditing AI agent decisions by keeping decision entries tied to confidence scores. This provides clear traceability for error-correction loops and helps evaluate the reliability of autonomous behaviors.

To add a new perception to an AI agent, you tell the agent: "After X, I now see Y differently" with a confidence score and an optional domain. This creates a perception shift that strengthens with reinforcement and decays when irrelevant.