Researchers propose a new AI agent behavior, Q&D, that learns to ask questions to retrieve required evidence and improve proactivity in task completion. They demonstrate its effectiveness on multi-hop question-answering benchmarks and in a simulated customer-service scenario.