The agent can finish it. That is not the decision.

The question

A working version used to be expensive, and the expense did some of the choosing. A weak idea often died while people argued about whether it deserved the weeks, before anyone wrote the code. AI-Assisted Product Discovery: When Code Is Cheap, Judgment Is the Job is Stefan Wolpers on what happens once that expense is gone. Agentic coding, an assistant that can carry an idea through to a working version, has removed much of the old gate. He says the scarce part moves to judgment: which idea deserves a customer's attention.

I liked the old gate more than I admitted. If the build would take the team a month, I did not have to be brave. I only had to be unwilling to spend the month.

The part I can check against my own week is narrower than his account. The weeks a feature used to cost were doing some of the refusing. An afternoon with an assistant no longer does.

The issue

The issue is a finished candidate that stays because it exists, while nobody has written why the other ideas wait.

Wolpers names two responses that look like opposites. One camp builds every idea the assistant can finish and lets the market sort them. The other wants the assistant to choose, and to attach a score. He says both hand off the part that just became scarce. Producing versions is now cheap. The customer still has to make sense of five half-finished features, and someone still has to maintain them. An assistant that chooses has taken the accountability with it. He shows the shape of a prepared brief on a reporting module I do not have. The checklist around that module stays in his article.

He says a language model tends to agree with the person asking. His countermeasure is two briefs, one arguing for the current goal and one arguing that the goal is wrong. I have not run those briefs, so I will not treat them as a result I have checked.

The person who loses is whoever has to keep the feature that was cheap to start, and the customer who received it before anyone decided it was the one worth having.

The refusal that can still win

On Orbit, the site where I publish these essays, an assistant may write the draft. It does not publish. I approve the one I will stand behind, and I can leave the draft unpublished. A draft that reads cleanly is still a candidate. Leaving it unpublished is a call I am willing to make.

That practice is smaller than the discovery check Wolpers describes. What I keep is the right to refuse a thing the assistant has already finished. In delivery work the change I sequence is the one already waiting, until I write why this candidate replaces it. A same-day draft does not take that place by arriving finished.

I have not counted how many ideas used to die from cost alone, so I cannot call the old gate a good filter. Wolpers notes that bad ideas still got built when a senior sponsor had the budget. Cost no longer does the refusing for me. If a finished draft is the choice, the build picked.

This fails when I have already started defending the draft and the refusal arrives as an afterthought. Then the assistant's first complete version is the thing I am protecting. I still let the draft exist before I decide. I have not checked whether a refusal written before any draft would change what I keep. Until I have, the rule stays small. The assistant prepares. It does not move the work onto the list of things I will stand behind.

Takeaway

A candidate an assistant can finish does not earn its place by being finished. I write the refusal where it can still win, and I keep only the one I will defend. If I have not made that call, the draft stays a draft.