Designing When AI Agents Should Step Back

Agentic AI can handle coding, research, travel planning, and even customer support, but success depends less on capability and more on how well human-AI coordination is designed. Simply building an "autonomous" system isn't enough—what users see, what they delegate, and what they control all need to be aligned together.

The authors break coordination down into three axes. Human involvement is the level of user engagement and monitoring, AI salience is how visible the AI is, and AI activity covers the actual behavior of the system, including invisible background work. Good UX means aligning these three axes to fit the task and context.

From this perspective, agentic experiences fall into three zones.

• Done with me: A collaborative experience where the user and AI work together across multiple steps
• Done for me: An automated experience where the user only initiates and reviews, while the AI handles most of the work
• Done under me: A supportive experience where the AI quietly helps in the background, largely unseen

The key point is that no single mode is the right answer. Even within the same product, different zones are needed depending on task difficulty, user expertise, and risk level, and trust and control are only maintained when the level of involvement and transparency is properly matched.

To explain this, the authors also propose a concept called the coordination curve. Involvement rises early in a task to set goals and constraints, drops during execution, and rises again during review and next steps—meaning human participation and AI salience move up and down over time. In particular, for long research tasks or multi-step workflows, there's a "valley" period where the AI works independently, making notifications, approvals, monitoring, and audit features important.

As a real-world example, they introduce an approach called responsive salience. The system continuously monitors signals such as task complexity, risk level, user expertise, and comfort, and when it judges that trust is low, it shows more detailed explanations, adds more approval steps, and increases transparency. Users can override this if they want, and when conditions change, salience quietly returns to its default level.

Early testing produced mixed reactions. Some users found high salience visually tiring, while others found it helpful because it suggested what to do next. Many participants didn't even notice the responsive salience itself because the system felt so natural—showing that well-tuned dynamic coordination can actually feel non-intrusive.

The conclusion is clear: it's not enough for agents to simply get smarter. They need a UX layer that adjusts when to collaborate and when to step back, based on the user's trust and expertise and the task's risk level. Only when that coordination is right do people come to trust and adopt agents—and even enjoy the experience of working alongside them.