Tag
This position paper argues that AI agents must be evaluated like behavioral systems through systematic observation and testing of their processes, not just performance outcomes, and proposes a research agenda to develop rigorous behavioral tests for better understanding and safety.