AI systems /
Why software is being rebuilt for AI agents
Most of the software my agents touch, they never actually open. They read the orders, pull the ad numbers, draft the campaign, answer the email. Not once does anything log into the dashboard the vendor spent years designing. That dashboard was supposed to be the product. It is starting to look like scaffolding.
We built software for hands
For about thirty years the interface was the product. Companies competed on the dashboard, the menu, the onboarding flow, the exact spot the button sits. All of it assumed one kind of operator. A person with a mouse, eyes on a screen, an afternoon to learn where everything lives.
That operator is changing for a lot of routine work. Not the person. The hands. The thing reading inventory and writing the first draft of the email is increasingly a model calling an API, not a human clicking through a UI.
A different operator wants different things
A person and an agent want almost opposite things from the same software.
- A person wants a clean screen, sensible defaults, and a button where they expect it.
- An agent wants a complete API, stable names, clear permissions, and data it can read in one pass.
- A person learns your interface once and forgives its quirks. An agent has no patience for a capability that only exists as a sequence of clicks.
- A person eyeballs a missing number and moves on. An agent needs the field to actually be there.
The software that is easy for an agent to drive is not always the software that looks good to a human. For years those were the same problem. They are starting to come apart.
The pace is not the point
It is easy to get lost in how fast this moves. A trick that mattered three months ago is already replaced by a better one, and that speed gets the headlines. It makes people feel they have to react to every week's news.
The direction is the part worth planning around, and the direction has been steady. The operator is changing. More of the work that used to need a person at a keyboard is being handed to a model with access to the same tools. You do not have to predict the timeline to plan for where it points.
Most roadmaps are built for the operator who is leaving
This is where teams get caught out. They are still planning like this is a normal cycle. The roadmap adds another dashboard view, another settings page, another report a human will read. All of it polishes the surface a person touches.
Meanwhile the operator that will actually drive the product is reading the API underneath, and that part is an afterthought. Undocumented. Half exposed. Locked behind a UI that assumes hands.
The teams that have noticed are rebuilding around the agent.
- They expose what the product does as clean tools, not just screens.
- They write documentation a model can read without a person translating it.
- They treat their API, and increasingly MCP, as a real surface, not a back door.
- They make their data legible, because an agent cannot eyeball what a person used to fill in by feel.
None of this kills the human interface. It moves the centre of gravity. The screen stops being the product and goes back to being one way in. I wrote about where that screen might end up in the interface after apps. This is the other half of it. Not what we look at, but who the software is built to be operated by.
What this changes for anyone building software
If you build software, the question quietly shifts. It used to be one thing. Is this easy and pleasant for a person to use. That still matters. But a newer question now sits next to it. Can an agent operate this without a human in the loop for every step.
That second question rewards boring things. A complete API. Honest permissions. Stable names that do not break when you redesign the screen. Documentation written for a reader who has read everything and remembers none of your meetings. Data that is actually filled in.
The point is not that interfaces stop mattering. The point is that the interface stops being the whole product, and the layer underneath, the part we treated as plumbing, becomes the thing you get judged on.
The rule
Stop judging your software only by how good the interface is. Ask whether it has a real API, clear permissions, and data a model can read without a person translating.
The interface is built for the operator you have today. The one arriving has no hands, reads everything, and never logs in. Build for that one too, and the dashboard nobody opens stops being a problem and starts being the point.