EvaluationsRealCLI measures whether an agent can still collaborate while it completes the task
The paired benchmark changes when information arrives, revealing a gap between execution scores and interaction quality under scripted user interventions.