Designing Non-Intrusive Proactive Interventions and Understanding Their Effects

Determine how to ensure that proactive interventions by always-on smart-glasses agents are perceived as welcome rather than intrusive, and ascertain how such proactive assistance affects user agency and decision-making over time.

Background

Proactive assistance could leverage ongoing egocentric sensing to surface timely suggestions (e.g., retrieving shopping lists in a store, offering meeting briefings). However, proactive prompts risk intrusiveness and may alter users’ sense of control.

The authors identify two concrete open questions: how to design proactive interventions that are welcome rather than intrusive, and how these interventions impact user agency and decision-making trajectories over extended use.

References

Important open questions include how to ensure proactive interventions are welcome rather than intrusive, and how proactive assistance affects user agency and decision-making over time.

— VisionClaw: Always-On AI Agents through Smart Glasses  (2604.03486 - Liu et al., 3 Apr 2026) in Section 8 Limitations and Future Work — Proactive Always-On AI Agents

We also assume uniform action costs, leaving a full model of cost and value, along with stronger methods for each subproblem, to future work.

— Robots That Take Initiative: A Framework for Building and Evaluating Proactive Robots  (2609.28910 - Patel et al., 24 Sep 2026) in Section Conclusion and Limitations

We leave a more involved consideration of the effect of robot actions on human goals to future work, as it requires studying human behavior adaptation in more detail.

— Robots That Take Initiative: A Framework for Building and Evaluating Proactive Robots  (2609.28910 - Patel et al., 24 Sep 2026) in Appendix, Section Misalignment between offline and closed-loop evaluation, subsection "Conclusion"

The discussions further highlighted that evaluation remains a central open challenge. Traditional metrics such as relevance, accuracy, engagement, and task completion capture only part of what matters for proactive agents.

— Report on The 1st Workshop on Human-Centered Proactive and Personalized Agents for Interactive Information Access at CHIIR 2026  (2608.18638 - Kaur et al., 19 Aug 2026) in Introduction; further discussed in Section 5, “Evaluation Beyond Accuracy”

We propose this as a hypothesis for future investigation: that unsolicited, low-investment suggestions may not only preserve autonomy but actively broaden the writer's idea space by reducing selection bias inherent in prompt-driven workflows.

— Towards Cognitive Process-Aware Proactive Writing Support  (2608.30424 - Yoshida et al., 31 Aug 2026) in Section 5, Discussion, subsection “Balancing Proactivity and User Autonomy”