When prototypes become effortless, we need to build judgment into the process, not just speed into the tools.
Agent coding has become remarkably good at answering one question, “does it work?” What I’m not seeing is a system for the other question, “should it exist?”
When a small team can generate hundreds of PRs a day that meet strict quality standards, the first question gets answered constantly. The way we decide what to build hasn’t caught up.
I’ve often been the person whose job it was to turn ideas into reality. An idea would get sketched on a whiteboard, refined through discussion and research, then planned and built. When that process failed, it was usually because the job was too big, the people weren’t available, or something else got prioritised. When it succeeded, there was usually something real in hand that broke through those headwinds: a prototype, a source of data, anything tangible that brought the idea to life and forced a change in plans.
Agents flipped that script. Any idea, of any size, can become tangible and nearly releasable in a day or so. The tangible thing now arrives first, before the idea has faced any of the scrutiny that used to come with it.
A running prototype feels like evidence. It has buttons and data, and it does things. But much of the time it’s really just an opinion that compiles. It gets built to win the room, not to learn anything, so what it mostly collects is agreement.
When my kids were small, they’d put their pants on backwards and proudly tell us, “look what I can do!” Dressing yourself is a real milestone. But we didn’t send them off to school like that.
Sometimes in demos someone shows a flow that took an afternoon to build and proudly announces, “look what AI can do!” Someone may challenge it with an alternative view, data about customers, or even a strategic flaw. With five simple words the debate is over, “We can change it later.”
Market research, customer interviews, design principles and domain expertise can all be sitting right there in the room and still fail to make it into what gets built, because the power sits with whoever holds the working artifact. Nobody decided that the customer’s voice should matter less. It just doesn’t arrive in the form the system rewards. A prototype arrives as a thing. Insight arrives as a conversation, a critique or a comment in a thread, and that’s easy to set aside when there’s a working thing on the screen.
Most of what goes wrong follows from that. Status drifts toward whoever can generate the most, the fastest. Discovery and design start to look like overhead. Screens get crowded because nothing forces anyone to choose. And each rushed decision gets layered on top of the last.
So what role should a prototype play now?
Not every prototype needs a hypothesis or a destination. Sometimes itches need to be scratched, and there’s nothing more satisfying than turning imagined possibilities into apparent realities. That’s why many of us are sleep-deprived these days…it can be hard to stop when you’re in the middle of an amazing creation.
And the prototype is obviously a great way to articulate a vision and get people behind it. The mistake is treating it as a blueprint. A prototype should make an idea easier to question by shedding light on problems we’ve ignored or new opportunities that have been invisible to us. If nobody can challenge the assumptions behind it, it isn’t a prototype anymore. It’s a pitch.
Customer research, design principles, analytics and domain expertise should all have something to say about what we’re seeing. A good prototype gives those perspectives a platform to speak from. It listens and reacts. A bad one simply collects applause.
If insight loses because of the form it arrives in, the fix is to give it a form that competes. Insight needs to compile too.
The harness developers already work in is the obvious place to do that. Right now it enforces code that passes. It could just as easily enforce product standards. A product review agent loaded with customer research, design principles and analytics could test the prototype’s assumptions against every screen and flow. Who is this for? What are they trying to do? What do we expect them to do differently, and what have we seen that says they will?
In practical terms, as an initial solution, the product reviewer agent could simply classify what kind of decision the evidence supports, such as “ready for limited release” or “user testing only” or “requires more evidence”.
The fair objection is that this becomes another gate, another way to make the process slower. And unlike a type checker, a product review agent can’t prove an idea is good. But it can make the assumptions visible, identify where evidence is missing, and insist that someone owns the decision to proceed anyway. That’s already an improvement over letting the demo make the decision for us.
It also shifts who has leverage. Once insight is built into the system rather than voiced in a meeting, “we can change it later” stops being an answer, because the team will expect the prototype to answer its questions before it moves forward, the same way code has to pass its tests. Whatever the answer, it forces an informed choice. We can still take risks, but we can’t pretend we didn’t know we were taking them.
Shipping a prototype without answering its questions is like sending a reporter’s raw notes out as a breaking news alert. The editor’s job applying context and understanding is crucial. They ask is it true? Does it matter? Who needs to hear it? The news org whose editor’s default is “we can change it later” will go bankrupt from legal fees long before their business model fails.
Maybe everyone in product development should spend a week on a news desk, watching how a relentless flow of new information shapes consequential decisions about what to put in front of users all day.
Leave a Reply