Make no mistakes
If soon anyone can build the software they need, why would we still need developers? Some thoughts on ambiguity, prototypes and pruning.
There’s a joke about coding agents: “Build a billion-dollar startup. Make no mistakes.”
It’s funny because it’s true. That one sentence leaves out everything that makes building software hard.
The agents themselves are impressive. Claude Code on ultracode can research a problem with many agents at once, derive requirements from what they find and build the features on its own. So the question is fair: if soon anyone can build the application they need, why would we still need developers?
I think the biggest unsolved problem for coding agents is ambiguity. Many of my experiments worked and still weren’t what I had in mind. They solved the problem, just not the way I wanted to solve it. Often that wasn’t the agent’s fault. It’s hard to describe what you want when you aren’t sure yourself, and most people don’t know what they want, or what a solution could look like, until they see one.
So I’ve stopped describing the answer up front. I let agents build a few versions quickly. Then I pick one, take the best parts of several, or at least find it much easier to say what I don’t like.
When I wanted new looks for the timer in Pomodoro Ledger, I had a gallery of eight built and liked exactly one. Most of the others were drawn objects: a tomato, a vinyl record, walnut. Seeing them told me what I wanted: screens, not objects. I couldn’t have written that into a prompt beforehand. For the Insights page, the gallery had 22 options. I kept eleven, and none of the three the agent recommended.
That’s how I think building will work for a while. It’s a decision tree: agents grow the branches, and a person prunes the directions that are wrong.
Right now the limit is usage. I can only run so many agents, so I have to focus on a few projects and choose a direction for each. Ideally, agents would run all the time: finish a task, suggest what could come next on their own, and let me decide where to go. At night I’d start the tasks that take longer, and in the morning I’d check what came back, prune, and start the next round.
So that’s where I see the value for now: in steering. Someone still has to hold the vision, look at five versions and say which one is right. And since most of us only know what we want once we see it, that someone needs a fast way to see it. That’s the part agents are already good at.