Showing what an AI agent is doing: plans, progress and results

Jakob Nielsen’s response-time limits give a rough guide. Up to about a second, people stay with the task. Up to about ten seconds, they keep their attention on the screen. Beyond that, they start doing something else and need to know when to come back.

Most agent tasks go well past ten seconds. A spinner and the word “Thinking…” tell someone that something is happening, but not what, how far it has got, or whether it has gone wrong. After a minute of it, people start to wonder whether to cancel.

Here is the same task shown two ways, three and a half minutes in.

For a task of more than a few steps, show what the agent intends to do before it starts. Keep it short, three to six steps, each a plain sentence about what the person gets: “Search flights from Gatwick on 14 March”.

Let people change the plan before it runs. Collect the questions the agent will need answered, such as which card to pay with, before it starts too, so the rest can run without anyone watching.

Mark the step the agent is on and say what it is doing, in the person’s terms. “Comparing 12 flights” helps. “Calling search_flights” doesn’t, however accurate it is.

Count real things. An agent rarely knows how long a task will take, so a percentage bar is usually a guess, and a bar stuck at 99 percent does more harm than no bar at all. “Checked 8 of 12 flights” is true, and it moves.

Show results as they arrive. When a step finishes, put its result under it in a line: “Found 12 direct flights under £200.” Someone who glances at the screen then sees progress, not just activity.

When the plan changes, say so. Agents change course for good reasons: a flight sells out, a page won’t load. A line such as “The 09:40 sold out, so I’m checking later flights” keeps the person’s picture of the task in step with what is really happening.

Keep the detail one tap away. Some people want to see every page the agent opened and every search it ran. Put that in an activity log under each step, not in the main view.

Here is a short task from start to finish: steps ticking off, a count rising, a change of plan, and the result.

Sometimes the agent can’t carry on without an answer: a choice between two flights, a missing detail, an approval to pay. Make that moment hard to miss. Put the question at the top of the task, change the task’s status to say it’s waiting, and if the person has left, send a notification that contains the question itself.

If the person can answer from the notification, even better. And if other parts of the task don’t depend on the answer, keep working on them while it waits.

Tell people plainly that they can close the window: “You can close this. I’ll let you know when it’s done.” Then make it true. The task should keep running on the server, not stop when the app goes into the background.

Send a notification when the task finishes or when it needs them, and say which: “Your Lisbon flights are ready to compare” is far more useful than “Task update”. If several tasks can run at once, give people one place to see them all, each with its status.

End with a report people can check quickly. Lead with the outcome, list each action with what it found or changed, and link each result to where it came from: the page, the email, the file. Say clearly what the agent couldn’t do and what is left for the person.

Keep a record afterwards. An activity log with the time of each action helps people check a task the next day, and helps your support team when someone asks what the agent did.

Screen reader users can’t see a step tick off. Announce changes in status, such as a finished step, a question or a finished task, through a live region, without moving focus. WCAG SC 4.1.3 Status Messages covers this. Keep the announcements short and don’t announce every line of the activity log, or the important ones get lost.