The Agent Does Not Get to Pick

The screen had one job. Tell an operator whether a synced contact list could be trusted.

Before anyone designed a pixel of it, the tool dealt four cards. Exception Feed. Master Detail Split. Entity First. And a visual direction: the Vignelli New York Subway Map.

The AI coding agent building the screen did not choose them. It could not swap them. It wrote down what it was dealt and built from there.

A subway map for a sync-status screen sounds like a joke. It is exactly the card a picker skips.

A shortlist collapses to its first option

A model asked for design directions gives you a varied shortlist. Then it picks the same winner, nearly every run. That is not a hunch. It is the stated reason, in a comment at the top of a script called the dice. The comment goes on: hand that shortlist to any selector, the model itself, a scoring pass, a simulated user, and the choice collapses back to option one. Only assignment survives.

I took the idea from Impeccable's published research, and the script's comment cites it. It went into the first commit.

In my own agent work, left to its defaults, a coding agent reached for the same component set every time. The result is competent and interchangeable, and it was never a decision. The usual fix is to ask for options and pick the best. That sounds like deliberation. It is the same failure one step removed, because the picker has seen every template and of course chooses the plausible one.

So the tool does not offer a menu. It deals a hand: three structure laws, which are how the surface is organized, how it handles depth, and how it frames the first thing you see. Then one visual direction. All of them come from fixed decks, and the agent records what it got.

The subway map had a reason

The structure decks hold 26 cards. Ten ways to organize a surface, six depth strategies, ten framings. Urgency Triage Board. Progressive Reveal. Cost Of Inaction. The direction deck holds 30 more, and none of them are styles. They are artifacts that real people shipped. A 1990s Bloomberg terminal screen. A Polish circus poster series. An Admiralty chart. Before a mode filters the decks, that is 18,000 possible hands.

Look at what that 1972 subway map actually does. It uses one decisive vocabulary for state, drawn the same way in every corner. It shows what the trains will do, not what the city looks like. That is a sync screen's whole job. The contract records the card as tone only. The brand's tokens stay and nothing gets repainted.

Nobody shortlists that card. Taste reaches for something calm and dashboard-shaped, and I would have too.

Same key, same hand

The sync screen's contract carries one line: exception-feed, master-detail-split, entity-first, vignelli-nyc-subway-1972, cns-6mwv78nd, generation 0. That is the whole hand, the key it came from, and how many times it had been redealt.

Run the same key and the same four cards come out. A bad hand in the field is a replayable bug report, not a mystery. Reroll, and the new hand is drawn from the same key with all four of those cards excluded. You cannot shuffle until you like the answer. You can only move forward into cards you have not seen.

An agent that may reroll freely will reroll until it gets the layout it was going to pick anyway. Chaining closes that door.

The hand can be wrong

Here is where it breaks.

An analytics surface was once dealt a fully typographic hand, with no room for charts. A rule in the tool says you keep what you are dealt. One of three blind judges dissented, wanting charts. A later round traced part of that dissent to a style-register mismatch, not the hand. The other two still ranked it first. The notes from that round say the argument structure, and not the roll, was what won.

A settings screen in another round came out as a consequence-ordered nav one run and a stack of form fields the next. The notes say the dealt hand dominates the outcome on that surface. These were informal rounds, not published records. They still say the same thing. Chance introduces variance, and variance is the price of breaking the gravity.

The tool has two answers when a card does not suit the surface. The first is that the direction serves the genre. On a trust surface, a payment form or a daily operating tool, convention owns the background, the chrome, the inputs. The dealt direction shows up in accents, type details and structure, and not as a costume. The rule says adapting is required fidelity, not a reroll.

The second answer is the review. A fresh reviewer, who never saw the build, scores whether the surface still serves its purpose. A hand that does not serve fails there. The dice do not decide whether a design is good. They decide where it starts.

A brief can forbid the default too

Human teams have the same gravity. Everyone has seen the default dashboard, and that is why it keeps getting built. A designer asked to pick a direction will often pick the familiar one and call it taste.

I think a brief can do the dice's work without the dice. Exception First is a card in the framing deck. As a brief it takes six words: lead with the exceptions, not the totals. That sentence has forbidden the default before anyone starts sketching. Cost Of Inaction and Objection First work the same way. The constraint is arbitrary, and it works, because arbitrary is what breaks the pull of the familiar.

It does not travel completely. People push back, and they should. A designer who hears the constraint and says it is wrong for this surface has done the job. An agent has no such reflex, which is why the constraint has to be dealt to it and not discussed.

I have also never isolated the roll. The evaluations compare the whole chain against a baseline. What I can claim is narrower. The dice stop the model from choosing, and choosing was the failure.

Twenty-two hands, no rerolls

In mid-October I searched the repos on my machine for every recorded hand. There were 22 distinct ones, from four projects. All of them were generation zero. Nobody had rerolled.

That might mean the cards were fine. It might mean a dealt hand creates its own momentum, once the contract has a name on it. The files do not separate the two, so 22 hands is a pattern, not a proof.

One surface in the same repos skipped the roll entirely. Its structure file says the hand was described and not assigned, and lists the limitation: taste-picked. Its file asks the reviewer to steelman a structurally different hand. That is the right instinct for every surface, rolled or not.

The roll does not make taste unnecessary. It moves taste to the one place it works, which is judging a hand after the fact, against evidence. Picking was never the part to hand a model.