GALAPAGOS · AUTUMN 2026JEJU · 33.4°N 126.5°E

Let's Build Agents, Whatever They Are

Why now is the time to build agents, in an era where our job has quietly become watching AI work.

Galapagos builds agents. Agents that begin from a user's intent and change real life.

Why agents, of all things, and why now?

1. People Began to Get in AI's Way

Last year, when Soyoung and I first built an app, we coded with Cursor. Even then AI was good enough that we could build an app without knowing frontend or backend. But I still had to spell out every detail myself. I was using AI, yet it felt like I was the one doing the work.

But in the spring of 2026, we realized something. Our job had become sitting still, watching AI work, waiting — and answering when it asked. Without noticing, the tables had turned. The thing getting in the way of the work was no longer AI, but us.

Until then, I had tried to put AI to use in my work. But now the question changed. "What environment do I need to set up so AI can work on its own?" AI handles things well by itself, so we just need to clear the road.

Once we saw this, we stopped grinding away for long hours at our desks. We sensed that everything was about to change. From then on, our conversations began with "agent" and ended with agent.

2. A Market Larger Than Any We've Seen Is Opening

Not every new technology creates a market and companies. Only a very few reshape an entire era. The internet did. Mobile did. In such eras the market is too large for a handful of companies to monopolize. Look at the App Store: countless apps, and each one is a company. Instagram is one app; so is Airbnb.

Until now, we saw AI only as deep-learning models and assumed a few big-tech companies would end up owning the market. It's true that the AI most people use is ChatGPT, Claude, or Gemini. But turn your eyes from the model to the whole picture, and you see a market too large for anyone to monopolize. The model companies matter, but across the whole market they are only one part — like the share a CPU maker holds in the entire computer and software market.

That whole market is the agent market. An agent is a service that begins from a user's intent and puts the period at the end of a task. Today, many people use AI as a machine to ask questions and get answers. But in some kinds of work — coding, for one — AI does more than answer. It becomes an actor in its own right: grasping intent and context, calling the tools it needs, planning, verifying, and finishing the job.

Building such an agent well is no small task. Because it doesn't move along a fixed path, you have to take security seriously, store conversations and context and pull them back out at the right moments, and put the right tools in its hands. And the better an agent is, the more deeply it digs into a specific task and context. So no single company can build every agent. Companies split off, as many as there are kinds of work.

As models advance, some companies will fade into history, but far more will grow alongside them. There could be as many agents as there are jobs in the world — perhaps even more. There is one thing the familiar law of supply and demand tends to miss: supply that creates its own demand. Economics has an old name for it — the Jevons paradox. In 1865, when steam engines that burned coal more efficiently appeared, coal consumption didn't fall; it exploded. As the cost dropped, the uses exploded. AI is no different. When something once unthinkable becomes possible, demand that never existed springs up. The more agents pour out, the more demand swells. A market larger than any we've seen may be opening.

3. The Language of Products Is Changing

Recently, Soyoung and I planned the UX/UI for a new product. As we always had, we listed the features we needed and drew the screens, tracing the flow of where you reach each one and where you find it again. We hadn't shipped, we had no users — and still we ended up with forty screens.

This happened because we assumed the user performs every operation. That is how it had always been for us. Even a simple transfer means writing down an account number, entering the amount, pressing the send button. All of it was ours to do. And so UX/UI evolved to be as simple and unconfusing as possible.

Now it's different. The user of an app's features is shifting from the person to the agent. An agent doesn't need a screen. What a person does at the screen now is not step-by-step input and operation, but conveying context. And even that can be conveyed by voice. Once that happens, most of those forty screens we drew no longer have a reason to exist.

Fewer screens is not the end of it. The very order — drawing screens first, then building to match them — changes. When and where to ask the user for intent; where to store past conversations and work and how to recall them; which directions things must never flow — these are the protocols we now have to design. Design still matters, but what we design is no longer the screen; it's the environment and the loop in which the agent works.

This is not the kind of change where a few screens disappear. The way we make things, and the way we use them, both change entirely. And no one can step around this shift. To lead the change rather than be dragged along by it, you have to jump in. That means building agents.

The permanent exhibition Designer Maker User at London's Design Museum includes this line: "Whatever form technological innovation takes, it is the work of designers and makers to turn its potential into change, enabling us to enjoy new technologies rather than be overwhelmed by them." In front of the new technology of AI, this is exactly what agents do: turning the potential of models into a form we can actually use without being overwhelmed.

It doesn't need to start grand. It's better to start from our own lives. Pick one thing you repeat every day, and build — with your own hands — a small loop and environment that can run it for you. Then explore for yourself how your days change.

In the end, the question narrows to two: What work should we put a period at the end of? And can we build an agent that does that work better than anyone? Galapagos is trying to answer these questions.