You don’t need to watch an agent while it fixes a banner.

In 2017, Attention Is All You Need introduced the Transformer, the architecture behind the models that have since changed the way we work. I’ll borrow the title to talk about a different kind of attention: yours.

Working with AI, I see more and more tasks that need a lot of it, and others that can already get zero attention while they run. Yet we still organize most of our work as if everything had to be watched.

1. Important doesn’t mean manual

A banner, a positioning bug, a responsiveness fix, a transition that looks off.

All of this affects the user, so it matters. How much it matters depends on the context. But in many of these tasks the decision has already been made. We know how the screen should behave. What’s left is making it work that way.

If someone could handle it for you and bring back the result for you to check, that would be gold. Well, with the models and harnesses we have today, that someone can be an agent.

It makes the change, starts the app, opens Chrome through MCP, tests different screen sizes, takes screenshots, compares the result with what was asked and fixes whatever is still wrong. It can even record a video as evidence. The whole loop of implementing, looking and adjusting can happen without you in the middle.

You define what needs to happen and you validate the delivery. You don’t need to pick the breakpoint, suggest every CSS property or keep staring at the terminal while the agent works.

It’s like hiring someone to paint a wall and standing behind them, inspecting every stroke of the roller. You need to agree on the color and check the finish. If you follow every move, you’ll spend the day busy with the painting anyway.

NOTE: a video shows what was tested, not everything that could go wrong. If the change touched a permission rule or hid an important action, the review needs to look at that too.

The importance of the result doesn’t demand your attention at every step.

2. “That’s a job for the frontend folks”

I’ve been seeing this on a team I work with. Before I joined, the flow was clearly split: one person does only frontend, another does only backend. And the team still gets stuck on trivial things because of it.

To fix a visual issue, you depend on someone else. You get in line behind whoever “does frontend”, even when the behavior is already decided and an agent can make and verify the change.

That split already bothered me. Today it makes even less sense. Why wait for the specialist to write CSS that the agent also writes?

Meanwhile, some questions do deserve attention: is this business rule right? Should this user be able to access this data? Did we understand the problem we’re trying to solve?

Some tasks have always required more judgment than others. Now you can hand many of the smaller ones to the agent, without tying up another person with them. But we keep distributing work the way we used to.

3. I wouldn’t take the courses I took

In The profession I studied for no longer exists, I told a bit about the start of my career.

During my technical high school program, HTML and CSS were among the first things I studied of my own accord. They were in the curriculum, but the content was very basic. I took Basic HTML, Advanced HTML and CSS courses from Fundação Bradesco. I also took a great course with Jeremy Foster and Michael Palermo on the now-defunct Microsoft Virtual Academy.

Today, I wouldn’t. And I don’t recommend you do, unless it’s a hobby.

As a professional investment, even for juniors, I’d spend that time differently. I’d work on real problems, with the agent, to learn to read the structure of a page, spot accessibility problems and investigate how the browser behaves. I wouldn’t go through a sequence of courses just to do that work by hand afterwards.

Akita wrote about this recently (in Portuguese): turning a finished visual decision into HTML and CSS is exactly the mechanical work we can already delegate. Choosing what the interface needs to communicate is still a separate task.

Karpathy recommended asking for explanations in HTML, with interactive pages and animations, to understand a subject better. The agent builds an interface that you use to learn and can throw away afterwards.

Look at the inversion: the thing I spent hours studying to produce became an output format I can ask for to understand something else.

I don’t regret those courses. They made sense at the time. But having learned that way doesn’t oblige me to recommend the same path to people starting now.

My attention is still limited. If I spend the afternoon following a responsiveness fix that the agent can solve and demonstrate, that afternoon doesn’t come back for understanding the product or digging into a hard problem.

Attention is all you need. Just choose better where to put yours.

Thanks for reading!