When Should AI Agents Ask? Clarification Requests Under Limited Human Attention

| Source: arXiv AI

Tags: Pairwise Coordination Policy, Multi-Agent Systems, Human-in-the-Loop, Dynamic Programming, Clarification Requests

A new model of multiple AI agents sharing one human's limited attention shows when each should ask for clarification, and its Pairwise Coordination Policy cuts cost 18-31% versus standard asking rules.

Details

When one person oversees several agents, each agent missing information can ask the human, search for the answer itself, or keep working on unaffected parts of its task. The authors model this as a dynamic program and show that whether to ask depends on more than an agent's own uncertainty, and that no per-agent ranking of questions is optimal. Their Pairwise Coordination Policy estimates how pairs of tasks delay each other. It is exactly optimal with at most two unresolved requirements and within twice the optimum in a broader case. On 864 synthetic cases its mean gap to optimal was 0.07%, cutting cost 18-31% versus three standard rules, and it had the lowest cost in an LLM agent case study.