Great roundup. One thing I'd add is that choosing an evaluation framework is only part of the process. The quality of the evaluation criteria matters just as much. Metrics like task completion, tool selection, trajectory quality, and consistency become much more reliable when combined with well-designed LLM-as-a-Judge workflows and human validation. We recently explored this topic at MLAI Digital, including practical guidance on building dependable LLM-as-a-Judge systems for production: https://www.mlaidigital.com/blogs/the-ultimate-guide-to-llm-as-a-judge-in-2026