Every AI product conversation eventually arrives at the same question: what does the interface look like? And almost every answer defaults to the same thing — a chat window. The user types something, the agent responds, the user types again. Repeat until the task is done or the user gives up.
That pattern is not wrong. It is just incomplete. Chat is one interface paradigm for agents, suited to tasks where the user needs to steer in real time and where the back-and-forth is part of the value. But many — probably most — agentic use cases do not work this way. The user does not want to converse with an agent over thirty turns. They want to hand off a task and get a result, with the ability to inspect, redirect, and approve at the points where human judgment genuinely matters.
The Handoff Problem
Designing for handoff is the core challenge of agentic UI. When do you interrupt the user? When do you proceed autonomously? When do you surface partial results for review before continuing? These are not UX questions — they are product decisions that depend on the trust model between the user and the agent, the stakes of the task, and the reversibility of intermediate steps.
Getting this wrong in either direction is expensive. An agent that interrupts constantly is just a slower, more annoying way to do the task manually. An agent that proceeds silently through consequential steps and gets something wrong has destroyed trust in a way that is very hard to rebuild. The product teams building in this space are learning that calibrating autonomy is as hard a problem as building the underlying agent capability.
Progress Visualization
Conventional software has a simple visual contract: button pressed, thing happened, confirmation shown. Agents break this contract. A non-trivial agent task might involve dozens of intermediate steps, tool calls, and decisions — none of which the user can see unless you build specific mechanisms to surface them.
The interesting pattern emerging from early production deployments is structured progress visualization — not a raw log of every function call, but a curated view of the meaningful decision points in the agent's work. Which tools did it use? What did it find? Where did it get stuck? Where did it make a judgment call that the user might want to revisit? That curated view is itself a design product, and the companies building agent platforms that get it right will have a significant advantage in enterprise adoption.
We are still early. Most agentic interfaces today are either raw chat or raw logs, with nothing in between. The design space in the middle is large and almost entirely unexplored. This is where I expect the most interesting product work to happen over the next two years.