14 design principles for human-agent interaction across 4 stages

A new paper from Haiyi Zhu, Canwen Wang, Qing Xiao, and Hong Shen argues that the real bottleneck for AI agents isn't technical capability—it's a lack of design knowledge for human-agent interaction. The authors introduce 14 design principles organized across four interaction stages: when the agent is first introduced, during ongoing interaction, as the relationship develops over time, and when failures occur. These principles aim to shift evaluation from pure autonomous task performance to interaction quality, trust, and adaptability. The team tested the principles against 9 existing agent systems, demonstrating how they provide actionable guidance for design teams.

Key principles include designing for transparency (agents should explain their actions), adaptability (agents should learn user preferences over time), graceful failure (agents should know when to escalate to humans), and appropriate autonomy (agents should not take actions unchecked). The research, submitted to arXiv as a position paper, emphasizes that agents are not just tools but collaborators—and their design must reflect that. For professionals building or deploying AI agents, this framework offers a systematic way to evaluate and improve user experience, potentially accelerating real-world adoption by making agents more predictable and trustworthy.