EASE: Efficient Human-Agent Communication under Evolving User Preferences

Agentic AI systems can execute increasingly complex real-world tasks, yet their transition from passive responders to active task coordinators raises a central communication challenge: how to efficiently incorporate necessary human input without overburdening the user. Existing systems often fail to adapt interaction forms and progress-reporting strategies as user preferences evolve, leading to redundant elicitation, over-automation, and verbose operational traces. In this paper, we propose EASE, an adaptive communication framework that enables granular control over user engagement, interaction form, and progress exposure. By mapping user preference stability, intervention signals, and task risk to a fine-grained automation spectrum, EASE elicits necessary user input with reduced user effort. To support timely intervention without cognitive overload, EASE surfaces only decision-relevant execution states while suppressing low-level operational details. Extensive experiments show that EASE consistently reduces time cost and cognitive load compared with strong baselines across no-history, stable-preference, and preference-drift settings. These results demonstrate that efficient human-agent communication is achieved not by maximizing autonomy alone, but by dynamically calibrating when, how, and to what extent users should be engaged.