The ETF Agent evaluation framework argues that a green test only confirms that a check passed under defined conditions, not that the user’s task was completed. It recommends separating runtime behavior, evidence retrieval, content judgment and user outcomes, while preserving failed, inconclusive and corrected observations so model and system quality can be compared on a valid denominator. The article combines practical guidance on RAG, traces, model judges, sub-agent contributions and cost with historical examples from its own development process.