I Stopped Writing Code. What Came Instead

Science fiction that became a workday

In the 2008 film Iron Man, Tony Stark has a home assistant called Jarvis.

It's a voice that runs the house, does calculations, checks blueprints, answers with a light touch of irony and understands its owner from half a word. In the film's novelization, the name is spelled out as Just A Rather Very Intelligent System Source 1.

When I watched the film, it was pure invention. A billionaire, his own lab, a system no one else has.

Now I can see that we're noticeably moving in that direction, and almost from home. You can consult AI, ask it to make a single part, a module, or an entire system based on your own idea. You don't need a lab. You need a laptop and an understanding of what you want.

To keep this from sounding abstract, let me describe what an ordinary day looks like for me.

My ordinary day

In the morning I open my plan and look at which task matters most right now. Then comes something I used to do rarely: I look at it from one level up.

The question isn't "how do I do this task" but "how do I build a system that will do tasks like this again and again." With clear rules, clear quality criteria and the right tools.

Take website development. First I break down what stages the work consists of and discuss with AI how to structure it. Then I gather the key information for the technical specification. We discuss the specification itself until it's sharp. And only then do I launch the team of agents.

This team has a lead agent, the orchestrator. It studies the specification, splits the work into parts and hands them out to subagents. Each subagent does its own piece of the work.

The easiest way to picture it is a construction site. The orchestrator is the foreman: he reads the plans, decides what to do first and hands out work to the crew. The subagents are the crew: each knows their trade and reports to the foreman. And I'm the client and the architect in one: I decide what we're building and by what rules.

A single orange wooden block on white, with five rows of white blocks fanning out from it — One foreman and five crews. That's how a team of agents works: the orchestrator splits the task, the subagents carry it out.

That's how the system I use to build websites works. I talk about it in more detail in What I Build.

It sounds smooth. In practice there's a place where everything is decided, and a place where things are still fragile.

What works and where it's fragile

AI already handles small tasks very well. I've stopped being surprised by how well. But with one condition: the task has to be stated clearly. I need to understand what result I want and what problem that result solves.

When that's clear, AI offers sensible solutions. We talk it through, settle on one option, and it implements it.

With complex systems the difference is even more striking. What I used to spend several months on, alone or with two or three programmers, now comes together in a couple of days. Systems with complex interfaces and well-thought-out architecture.

This is my experience on my tasks, not a promise. And it comes with an important caveat: I lay down the architecture. I set the frame and the rules: how to organize the code, what it should be like so that it can be extended and changed. Without that frame, a team of agents builds fast, but not necessarily something you can keep developing.

There's also a place where, honestly, I'm not sure yet. To build a system for a recurring task, you need experience solving exactly those kinds of tasks with AI. There isn't much of it yet. So some systems are hard to put together: it's not yet clear how exactly to organize them. I'm figuring it out as I go.

Before: I write the code myself.

Now: I set the task: what result I need and why.
Before: Me and two or three programmers.

Now: The work is done by an orchestrator and subagents.
Before: A complex system takes several months.

Now: One of similar complexity takes a couple of days.
Before: I think about how to implement it.

Now: I think about what I want to get and what problem it solves.
Before: I'm surprised by every good answer.

Now: Instead of being surprised, I set the frame and check.

From "how" to "what"

The biggest change happened in my head.

A programmer's work is mostly "how": how to structure the data, how to split the program into parts, how to get around a bug. Attention goes into implementation details, and there's not much energy left for the question "what do I actually want to get."

Now the proportion has flipped. I no longer write code myself; I trust AI with it. My work is the idea, the desired result and the frame within which it gets built.

My job now isn't how to do it, but what I want to get.

This didn't happen in a day. The shift is gradual, and I notice it in small things: in the questions I ask myself in the morning, in how differently I relate to the development process itself.

I think this shift is coming not only for programmers. Anyone who works with AI sooner or later has to answer the question "what do I really want." And that question turns out to be harder than "how do I do it."

And then it gets more interesting: what happens when AI starts not only writing code but also assembling the things around us.

Jarvis steps out of the laptop

I often think about how to connect devices to my laptop that would assemble, build, move and solder things. I don't need that yet. But I can see these possibilities becoming more accessible every year.

There's already evidence. On social media I see ordinary people at home designing real devices with ChatGPT or other models. The AI suggests which parts to buy, writes the program and helps load it onto the device. The person prints the case themselves. The result is a thing that does exactly what they had in mind.

These home-built projects are based on inexpensive boards like the Raspberry Pi and a 3D printer. A 3D printer is a machine that, following a design file, lays down an object layer by layer from melted plastic. Such a project used to require knowing electronics, programming and modeling. Now a significant part of that knowledge can be picked up in a conversation.

A light wooden table by a window: a small bare computer board, an open white 3D-printed case, a screwdriver with an orange handle and a few screws — A board, a case, a screwdriver. A home assistant people build themselves, to their own design.

From here it's easy to imagine a home where Jarvis is no longer fiction. You can talk to a voice assistant: it opens the curtains, turns on the kettle, tells you what's in the fridge and orders the groceries you buy regularly on its own. The lights come on just the way you need them right now.

And it's a system everyone builds for themselves, not an off-the-shelf "smart home" in a box. I think that's what sets what's happening apart from the science fiction in the film: people build the assistant themselves, without a billionaire's lab.

What would you come up with if you had someone to talk to who remembers everything?

The limit now is imagination

If AI can design a device, write a program for it and tell you how to assemble it, what's left for the human? I think the main thing: coming up with what's worth making.

There's now enough knowledge and enough hands. Everything comes down to how broadly you can imagine what your life and the environment around you could be like.

I see three directions worth putting the freed-up energy into. The first is basic needs: food, housing, everyday chores, everything that takes up attention every day. The second is access to information, when any question can be worked through based on data and research. The third is the environment around us: devices that take small, repetitive tasks off our hands.

They have one thing in common. Tasks that come up all the time should be solved quickly and without spending cognitive resources, that reserve of attention and energy that is finite for each of us.

That's exactly why I write on this site. I want to widen the field of imagination. To show that with these technologies life can be made noticeably better, longer, and freer for bigger things: for research, for caring for the planet, for people working together, for exchanging and creating resources. In a way where everyone wins.

Jarvis is no longer science fiction: some of his abilities are already available at home, with a laptop and inexpensive boards.
Working with AI is teamwork: the orchestrator splits the task, the subagents carry it out, the human sets the goal and the frame.
AI works well where the task is stated clearly: the result is understood, and so is the problem it solves.
For now, the architecture and the rules are set by the human. Without them, you quickly get something that's hard to develop further.
The main shift is in thinking: from "how do I do it" to "what do I want to get."

Who I'm looking for

I'm interested in people who are already building something with AI: systems, devices, ways of working. And in those who only want to start and are wondering where.

It's especially interesting if you have a project that could use both development and agent teams. Or just a question you'd like to discuss with someone who's walking this path right now. Let's meet.

Sources

Peter David, "Iron Man," the novelization of Jon Favreau's film Iron Man, 2008. The source of the J.A.R.V.I.S. acronym. Back to text
"$35 Raspberry Pi Model B Sold Out Within Minutes," Tom's Hardware, February 2012. On the launch of sales of the first Raspberry Pi model. Back to text