Sometimes AI feels like it can read your mind—until it doesn’t. Small gaps in what you meant (“cheap, but not a red-eye”) or what the AI assumed (“any cheap flight is fine”) turn into big problems once agents can act: editing files, booking trips, or moving money. If an agent misunderstands you—or gets tricked by a malicious web page (prompt injection)—it can buy the wrong ticket, leak info, or mess up your stuff. The risk isn’t just bad answers; it’s unwanted actions.

This talk turns the research on Challenges in Human-Agent Communication into a simple, practical checklist you can use right away. We’ll focus on four clear questions for every agent task:
Can-do (What are its abilities and limits?) → Plan (What will it do next?) → Doing (What is it changing right now?) → Done (What actually happened, with proof?). You’ll see concrete patterns like: show capabilities up front, preview with a diff before committing, use scoped permissions and approvals, attach evidence for results, and ask for help when uncertain.