Feedback Loops Skill

Purpose

Establishes a production-grade self-correction and verification framework for AI agent execution. Feedback loops operate on the principle that agent outputs must be systematically evaluated, validated, and corrected through structured cycles. This system implements Implement-Verify-Fix (IVF) cycles, agent self-reflection patterns, human-in-the-loop checkpoints, multi-layer output verification, automated validation hooks, correction trigger mechanisms, quality gates, and continuous improvement loops. The goal is to catch errors post-execution, enable iterative refinement, and maintain output quality throughout extended agent sessions.

Core Principles

1. Verify Every Output: No agent output is trusted by default. Every generated artifact must pass through at least one verification layer before being considered complete.
2. Structured Correction Cycles: When verification fails, corrections follow a disciplined IVF cycle. No ad-hoc patching or blind retries.
3. Human-in-the-Loop at Critical Junctures: Irreversible actions, high-risk modifications, and ambiguous decisions must route through HITL checkpoints.
4. Multi-Layer Defense: Stack multiple verification layers (syntax, semantics, integration, acceptance) to catch errors at appropriate abstraction levels.
5. Continuous Learning from Failures: Every correction event generates a learning signal. Track failure patterns to prevent recurrence across sessions.

Agent Protocol