Show the work, keep the magic

In Omaru’s onboarding, we ask you a handful of ordinary questions. While you answer, the product is looking you up in the background, learning more about you than you are telling it. By the time you are done, it shows you a picture of yourself that is fuller than the one you gave. That reveal is the moment the whole thing is built around.

I held something back on purpose so it would land. The instinct is old. Storytelling and magic shows have always chosen their timing. What is new is what I was holding back. It was the model’s output, live and a little different every time.

That is the shift I keep circling. For most of software’s history, what appeared on screen was decided in advance, a screen, a layout, a flow you could lay out and be sure of. Now the thing at the center moves. The agent produces something you cannot fully predict, and the decision grows to fit that. The real question becomes what to do, moment to moment, with output that is alive.

You cannot hold everything back

An agent that works in silence feels broken. If you cannot see anything happening, you assume it stalled, even when it is doing everything right. So the question is which part to hide and which to show. You hold back the mechanism, the how, and you surface the life, the fact that something is happening. Those are one decision seen from two sides.

Showing the life is three things

And showing the life turned out to be three different things. Blurring them is where it goes wrong.

There is the thinking stream, the tokens scrolling past as the agent works. It comes too fast to read, so it is not information. It is a heartbeat. It says the thing is alive, and it happens to be mesmerizing to watch, so it earns its keep twice, as a sign of life and as spectacle.

There is the trail of actions, the tool calls. Those are discrete and slow enough to read, so this is the honest observability layer, the place you can see what the agent actually did.

And there is the result, the reveal everything was setting up. Each of the three is its own decision, doing its own job.

More can be more

This is also why, with agents, more can be better. In Omaru, when several are working, you see them as cursors moving around the desktop at once, each its own color. You are not meant to follow any single one. That is the point. A screen full of agents at work is its own kind of spectacle, and while it entertains it says something true, that each one is operating a computer the way a person would.

What you can actually stage

There is a limit to how much of this you can stage. The flows you own, like onboarding, you can choreograph, you control the order things happen in. Once an agent is off doing open-ended work, the order is no longer yours, and the design becomes a standing choice about what to surface and how loudly, more than a script you write once.

All of this is design that only works if you know what the agent is doing underneath. When it is searching, when it is stuck, what it is about to produce. Which is why, building these things, designing them and building them have stopped feeling like two separate jobs to me.