The original title is about designing the moment when an AI agent needs human input. Let me rewrite this as a punchy headline.

Short summary

This article proposes a structured design pattern for the moment an AI agent must pause and request human input, treating it as a task-state handoff rather than a chat message. It covers checkpoint preservation, notification design, risk-tiered interaction types, expiry rules, and audit trails so users can resume work without reconstructing context. The author also outlines a five-scenario usability testing protocol and discloses a connection to the MonkeyCode project, clarifying the proposal is conceptual rather than implemented.

• •Frame agent questions as task-state handoffs with checkpoints, evidence, consequence, expiry, and recovery fields
• •Tier interactions by risk: clarification, approval, confirmation, and credential requests each need distinct UI treatment
• •Test with five concrete scenarios covering desktop, mobile, unauthorized access, stale context, and post-resume failure

Generated with AI, which can make mistakes.