Many open questions can be answered before building

The obvious objection to deeper discovery goes like this: we could simply build a lot and learn in production what actually works. Agentic development makes this tempting. If building is cheap, why not build everything you can think of and try it out?

This objection assumes that all relevant uncertainty sits at the solution level: Which solution works? But much of the uncertainty lies earlier, in the opportunity space. Is the problem real? Do enough people actually feel this pain? Is it intense enough that they would do something about it?

No feature, however quickly generated, will answer these questions. You can generate a solution agentically in a few hours and know just as little about the underlying opportunity as before. You have demonstrated feasibility, but not whether the opportunity exists.

The obvious counter-objection: in production we will see whether the feature is used, and that tells us whether our desirability assumptions hold. But usage metrics are far too fuzzy a signal for that. If a feature isn’t used, that could mean a lack of desirability, but also poor usability, poor discoverability, or users unwilling to give up a workaround they have grown used to. If it is used, that could be curiosity, a single power user, or use for an entirely unintended purpose. You end up with a confounded signal and have to work backwards to the actual cause. A short conversation before building would have answered the question more directly and more cheaply.

Whether an opportunity exists and how much it matters can be established through continuous interviewing and by comparing opportunities, not by testing solutions. Even at the solution level, desirability can often be settled with a conversation, an observation, or a fake door test. That takes hours, not sprints. Skip it and go straight to building, and you validate the wrong things, just faster.

Usability and feasibility assumptions, on the other hand, can often be validated more cheaply by prototyping or simply building than by interviews. Agentic development lowers that threshold further. Fast prototypes get cheaper, and the cycle between building and learning gets shorter. That shift is real.