Agent Evaluation Shifts to Real-World Quality, Adaptation and Robust Systems

Evaluation is moving beyond task success toward collaboration, procedural compliance, multimodal work, and realistic environments. RealCLI, Co-Design Bench, and instruction-capacity results report substantial gaps; DynaFlow claims large multi-agent gains through explicit information passing, but benchmark setup, a GPQA-Diamond regression, and reproducibility warrant scrutiny. New benchmarks broaden the landscape, while independent validation and real-world transfer remain essential.