evaluation

What problem does it solve? Agent systems are non-deterministic, take multiple valid paths to goals, and lack single correct answers, making traditional software testing approaches ineffective. This Skill provides structured evaluation methods that judge outcomes rather than specific execution steps, enabling teams to catch regressions, validate context engineering choices, and measure quality over time. ## Core Features & Use Cases - Multi-Dimensional Rubrics: Score agents across factual accuracy, completeness, citation accuracy, source quality, and tool efficiency with weighted 0.0-1.0 scales and configurable pass thresholds. - LLM-as-Judge and Human Evaluation: Combine scalable automated judgment with human review for edge cases, hallucinations, and subtle biases that automation misses. - Complexity-Stratified Test Sets: Build test sets spanning simple single-tool queries through very complex multi-step reasoning tasks, sampled from real usage patterns. - Use Case: A team shipping a research agent builds a test set of 50 queries across four complexity levels, runs LLM-as-judge scoring on every prompt change, and tracks weighted quality scores in a dashboard to detect regressions before deployment. ## Quick Start Ask the agent to design an evaluation rubric and test set for your agent system, covering multiple quality dimensions and complexity levels.