Agentic Eval

Implements self-critique loops where Claude generates output, evaluates it against your criteria, then refines based on its own feedback. Includes evaluator-optimizer patterns, test-driven code refinement, and LLM-as-judge scoring with JSON-structured critiques. Most useful for quality-critical tasks like code generation, reports, or analysis where you have clear success metrics. The reflection patterns prevent single-shot mediocrity by forcing iterative improvement, though you'll want to set iteration limits to avoid endless loops. Works best when your evaluation criteria are specific rather than subjective.