A team of researchers built a formal model of how AI agents should decide when to ask a human a clarifying question versus press ahead on their own. The core finding: asking has a hidden cost. Each question pulls the human's attention away from other tasks running at the same time. An agent that asks too often adds more coordination burden than the work it saves. The model quantifies this trade-off and gives agents decision rules to follow. The research treats human attention as a scarce resource shared across concurrent agent tasks. An agent that holds uncertainty should weigh the expected cost of acting on a wrong assumption against the cost of interrupting. The right threshold for asking shifts depending on how many tasks are running, how reversible the action is, and how confident the agent already is. A single-task scenario tolerates more questions. A crowded parallel workload calls for tighter question filters and more willingness to act without confirmation. For teams building agentic products, the practical takeaway is that interruption design cannot be optimized task by task. Queuing, batching, and timing of clarification requests need to account for what else the human is managing. The research also surfaces a failure mode: an agent calibrated for a light workload will over-ask when the human is busy. An agent calibrated for a heavy workload will under-ask when the human has headroom. The model suggests that agents should read workload signals before deciding whether to surface a question.