We stopped doing human-centered design

We have been steering away from human-centered design, toward something we have started calling agent-centered design.

Which is an odd sentence to publish. Human-centered design is not a trend you grow out of. It is the closest thing software has to a settled moral position, and moving off it sounds like the last thing a company says before shipping something nobody can use.

So here is the reasoning, in order.

Human-centered design was right about who does the work

For fifty years, software was built for humans to get the work done. The machine held the data, the person made the decisions, and the software sat between them, handing a person what they needed in a form they could act on.

Increasingly that is not what we ship. Our software is used by agents to get the work done. The human sets the goal and owns the outcome. The file work is done by something that has never seen a screen.

At that point you are not building an application. You are building a harness, and a harness built to the wrong specification is worse than none.

Design is optimization around the shape of the operator

Human-centered design is not a set of aesthetic preferences. It is optimization around one specific creature with specific strengths and specific weaknesses.

A person holds about seven things in working memory. They read top to bottom, they tire, they lose the thread on the third scroll. They are also extraordinary at judgment, at taste, at knowing something is wrong before they can say why.

So we built conventions that protect the weaknesses and lean on the strengths. Progressive disclosure, because attention runs out. Ranked hierarchy, because reading is linear. Dashboards, because a shape is faster than four thousand rows. Every one is a workaround for a human constraint, and every one is good craft.

An agent has none of those constraints. It also has none of those gifts. It is a different creature, with a different shape, and it needs a different environment.

“

For fifty years we optimized systems around the shape of a human. We are now optimizing them around the shape of an agent.

Take Amazon

Amazon holds, for any product, something like a hundred attributes it could let you filter by. Weight, materials, warranty length, seller tenure, return rate, ship-from country. Building all hundred into the sidebar would take an afternoon.

You get nine.

Not cost, not laziness. Somebody worked out which nine a shopper can hold in their head at once and threw the other ninety-one away. That act of subtraction is one of the more valuable things the company does.

Now notice what it is subtraction for. Those nine exist because of human working memory. They are a compression of the real system, tuned to a bottleneck, and the loss is the value.

Point an agent at that sidebar and you have handed it the compressed copy of your own system. It would have held all hundred attributes, combined them in ways no shopper would, and compared ten thousand records without needing a pleasant dropdown.

Every dollar you spent making that interface excellent bought you a ceiling.

Before you rip out your UI

The lazy version of this argument is that agents are humans with the limits removed. That version will cost you a quarter.

An agent reads a thousand records without going cross-eyed. It still falls over when:

• Two tools have overlapping descriptions and it cannot tell which one it wants
• A record has no provenance, so it cannot judge whether to trust it
• The context it needs exists, but sits where retrieval will never surface it
• State vanishes between runs, so it solves the same problem for the fortieth time
• It cannot tell a quiet success from a quiet failure
• Permissions are vague, so it either overreaches or stalls
• You gave it a workflow but not the primitive it needs when reality steps off the path

Not one of those turns up in a usability test. They are design failures, just not the kind anybody trained for.

What actually changes

Built for a person	Built for an agent
Features	Capabilities
Screens	Tools
Navigation	Discovery
Information hierarchy	Information topology
Dashboards	Structured state
Ease of use	Ease of reasoning

A feature is a capability plus the wrapping that makes it usable by a person. The agent wants it unwrapped.

The fourth row is the one that keeps surprising people. Hierarchy answers "what should I look at first," which is a human question, because attention is scarce and linear. An agent never asks it. Its question is "how is this connected to that." Rank is not the useful shape. The graph is.

Build primitives, not screens

Instead of a screen that tells the agent how to do the task, the harness hands it raw materials and lets it work out its own path.

Knowledge. Searchable, traversable, provenance attached.

Topology. A machine-readable map of where things live and how they relate.

Tools. Small, explicit, defined inputs and outputs. No two overlapping.

State. What happened, what exists now, what is still unresolved.

Memory. What it learned on file four hundred, waiting for it on file four hundred and one.

Goals. A definition of done, not a path.

World model. Enough about the environment and the rules to reason about consequences.

Observability. Whether the action landed, and how to recover when it did not.

Notice that none of it renders. That is rather the point. Our knowledge layer is written to be read by an agent first and a person second, and primary source verification is a capability our agents call, not a page a coordinator visits.

The shortcut that looks clever

If you already have a good product, the tempting move is to sit the agent in the human's chair. Point it at your own screens and let it click.

It demos beautifully. It is also the most expensive way to work ever devised. You render structured data into pixels, pay an agent to parse those pixels back into the data you already had, then cap it at nine filters.

One question for your product

Walk it, part by part, and ask: is this here because the work requires it, or because a person was going to be looking at it?

The second pile is not waste. People still need to see, approve, and put their name on the outcome. But it is no longer the system. It is the window into the system, and it is worth knowing which of the two you have been funding.