The Human-in-the-Permission-Loop: Adaptive Permission Models Guided by Real-Time User Feedback

TL;DR: What if permissions in agentic AI systems weren’t static or fixed, but adapted in real-time based on user reactions and feedback? Let’s test a system where users can “shape” permission boundaries as they interact with the agent.

Research Question: Can integrating continuous user feedback into AI agent permission systems lead to more trustworthy and context-appropriate security models?

Hypothesis: Dynamic, user-adaptive permission models will better align with individual preferences and situational needs, reducing both over-permissiveness and unnecessary friction compared to static permission modes.

Experiment Plan: Design an extension to Claude Code’s permission system allowing for real-time user overrides, confirmations, and feedback collection. Develop ML models that learn and adapt permission boundaries based on this feedback. Deploy the system in both simulated and real user environments, logging permission changes, override frequencies, and security incidents. Evaluate user trust, perceived safety, and operational efficiency.

References:

• Liu, J., Zhao, X., Shang, X., & Shen, Z. (2026). Dive into Claude Code: The Design Space of Today's and Future AI Agent Systems.
• Couch, C. W., & Miller, M. E. (2025). Ethical Human‐AI Agent Interface Considerations. INCOSE International Symposium.