Why I build agents that execute

A chatbot gives you advice. An agent does the work. I build the second kind.

Advice was never my bottleneck. Making the thing was. That is the part I hand to agents now.

The bottleneck moved

For most of my career, production was the constraint. You knew what to build. Building it took the week.

That is no longer true. A screen takes minutes. Forty screens take an afternoon.

I used to spend days producing what I now spec in an hour. The making is not where the time goes anymore.

The old advantage was being fast at production. That advantage is gone. Speed is a commodity now, handed to everyone at once.

When making things gets cheap, making the right thing gets expensive. The scarce work moved upstream of the file.

Advice is the easy half

Most AI products stop at advice. They tell you what a good design system looks like. Then they leave you to build it.

That is the easy half. Anyone can describe the target. Reaching it is the labor.

A chatbot will list the twelve traits of a strong onboarding flow. I have read that list. I did not need it written again.

I needed the flow, built to my system, ready to test. An advisor is useless once you already know the answer.

So I do not build advisors. I build agents that read my design system and act on it. They produce the files, not a paragraph about the files.

What I stopped doing by hand

I rebuilt this whole site through Claude Code. Every page is real HTML now. No JavaScript bundle, readable by people and machines.

I did not hand-write the boilerplate. I automated it. I wrote the intent and let the agent produce the pages.

I stopped writing the same token in eight platform formats by hand. I give the agent the source of truth. It produces the variants and I check them.

The tedium was never the thinking. It was the typing. I handed the typing over and kept the thinking.

None of that was design. It was assembly. Assembly is the first thing I give away.

I run my whole operation this way. Chat for thinking. Code for building. Cowork for orchestrating. It is not a tool I open sometimes. It is how I work.

From doing to directing

I am Head of User Experience and Design at 10Pearls. Leading a team taught me to direct instead of do.

Working with agents needed the same instinct. One agent does a task. Cowork runs several at once, and I steer.

I moved from making one thing well to directing many things at once. That is a different muscle. I had already trained it on people.

The best agents feel like good hires. You brief them well, then get out of the way. A vague brief produces vague work, from a person or a model.

The scarce thing

You can make forty things this week. You could not do that two years ago.

The question is which of the forty is worth making. Nothing autocompletes that answer.

Say I can build ten onboarding flows by lunch. The build is free now. Knowing which flow a real person will finish is the whole job.

The scarce assets are curation and judgment. That is where I spend the time the agents give back.

I teach my design team the same shift. Think in decisions, not screens. The screen is now the cheap part.

The system agents read

An agent cannot ship against a design system it has to guess at. Give it a vague one and it guesses teal, picks the wrong border, invents a shadow nobody approved.

So I built one it can read. skill-kit is a 577-file design knowledge graph. Nine categories, distilled from 24 platforms. It is public, MIT, at github.com/Youmair/skill-kit.

With skill-kit in context, an agent reads a component and builds a screen that already matches. No back-and-forth. No correcting its guesses one comment at a time.

I scaffolded far more than that. Then I chose to cut it down to 577. The collection was the easy part. Deciding what to leave out was the work, and I wrote up how I did it.

The point was never the count. An agent reads it and ships, instead of asking me what teal means.

The other half

I wrote before that AI is not replacing designers. Read that here. This is the other half.

That piece was about what stays human. The judgment, the translation, the decision. This one is about what I hand to the machine. The production, the scaffolding, the tedium.

Both are true at once. I give the machine everything it can do. I keep the part only I can.

The take

I do not want an agent that advises me. I have opinions of my own.

I want one that executes, so I can spend my hours deciding what is worth executing.

Production used to be the bottleneck. Now the bottleneck is knowing what to build. That is a better problem, and it is the one I work on.