Can an agent complete this flow?

For many years, we've been designing flows for people. Now, I think we need to start designing them for agents too.

Lately, I've been thinking about a simple question:

Can an agent enter your product, book a meeting, order a service, complete a form, or perform another key action without getting stuck?

Last week, I watched an agent go through one of our prototype flows.

It didn't see the screen. It read the accessibility tree: roles, names, and states. An icon button without a label wasn't a button to it. A success toast without aria-live never happened. It submitted the form twice.

And none of these were bugs. Every human test had passed the flow. And that's what I find interesting.

We already think about user flows. I think we're going to need to start thinking about agentic flows too.

Accessibility used to be seen as a compliance task at the end of a sprint. But now it can also become the primary interface for a new type of user: an agent that can book meetings, register attendees, fill out forms, and complete tasks on behalf of a person.

And it turns out that designing for this kind of user is very similar to the work we should have been doing anyway: making interfaces accessible, semantic, and unambiguous not only for people, but also for machines interpreting them.

So I'm experimenting with a new artifact alongside the user flow: agent walk.

We give an agent a task without additional instructions and record what it perceives at each step and where it has to guess.

Essentially, it's a usability test where the participant is reading a completely different version of the interface than the one we designed for people. I think this will become a basic product requirement sooner than we expect.

Soon, we might not just ask: "Can a user complete this flow?"

But also: "Can an agent?"

If you've done anything similar with your product, I'd love to compare notes.

Have you tried anything like this?