Taste Is a Bottleneck: Where Judgment Has No Substitute Yet

Ask an agent for ten approaches to a problem and you'll get ten reasonable approaches. Ask which one to take and you'll get a well-argued pick. Watch experienced engineers evaluate the same ten and something different happens: they discard seven almost immediately, on grounds they'd struggle to articulate, and are usually right to.

That fast discarding is what people mean by taste, and it's the input in shortest supply as generating options gets cheap.

What taste actually consists of

Not mystique. Three components, all of which are compressed experience:

Pattern recognition over outcomes. Having seen this shape of solution go badly before. Not a rule — a recognition that produces unease before it produces an argument. The articulation comes afterward, if at all, and its absence doesn't make the judgment less reliable.

Sensitivity to what's load-bearing. Knowing which parts of a design will be hard to change later, and therefore where to spend care. This is what separates the person who over-engineers uniformly from the one who engineers heavily in three places and lightly everywhere else.

A model of the future. Where this system is heading, what will be asked of it, which flexibility will be needed and which will be dead weight. Grounded in knowing the domain and the organization, not in general reasoning.

Why generated options make it more valuable, not less

The intuition is that if a model can produce good options, taste matters less. The opposite happens.

When producing an option was expensive, you evaluated two or three. Now you can have twenty, all plausible, all internally coherent, all defensible in isolation. The scarce step is no longer generation — it's discrimination. And discriminating between twenty plausible options requires more taste than choosing between three, not less.

⚠️ There's a specific hazard here. A well-argued case for a mediocre option is more persuasive than a poorly-argued case for a good one, and agents produce uniformly well-argued cases. Someone without independent judgment will be moved by argument quality, which is uncorrelated with option quality. That's a new failure mode, and it lands hardest on people who haven't developed the discrimination to resist it.

Can it be transferred?

Partly, and the parts are worth knowing.

What transfers: explicit heuristics, once articulated. "Don't add a configuration option unless two customers have asked" is teachable and encodable. A lot of what looks like taste is actually rules that nobody wrote down, and writing them down works.

What doesn't transfer easily: the recognition that fires before articulation. Someone can tell you the rule they extracted; they usually can't tell you the fifty cases that produced it, and the rule alone is more brittle than the recognition.

What can be approximated: an agent given a written record of past decisions and their outcomes has something closer to taste-informed input than one given none. → This is a strong argument for keeping decision records — not for process compliance, but because they're the only externalized form of the thing.

💡 How it's developed

Uncomfortably, by doing the work and seeing how it turns out — which is the activity being delegated.

Taste comes from having built things, made choices, and lived with the consequences. Reading about a decision doesn't produce it. Reviewing someone else's decision produces some, weakly. Making the decision and being wrong produces a lot.

This is the same tension that appears around reviewing agent output: the ability needed to direct the work is developed by doing the work. There's no clean resolution. The practical version:

• Make some decisions the slow way, deliberately, and follow up on how they turned out. The follow-up is the part that converts experience into judgment, and it's the part almost everyone skips.
• Keep a decision record — what you chose, why, what you expected, what happened. This is how you learn from your own decisions rather than merely accumulating them.
• Stay close to consequences. Judgment develops where you're around long enough to see the results. People who move on before the consequences arrive don't develop it, regardless of how many decisions they made.

✅ Using it well right now

• Generate broadly, discriminate narrowly. Have the agent produce many options; spend your attention on cutting rather than on producing.
• Discard on your own grounds, not on the argument's quality. If something feels wrong and you can't say why, that's data. Investigate it rather than deferring to the better-written case.
• Write down the heuristics you notice yourself using. Those are the transferable part, useful to your team and to your agents.
• Protect the conditions that build it — hands-on work, ownership, and being present for outcomes.

The takeaway

Taste is compressed experience: recognizing shapes that went badly, knowing what's load-bearing, holding a model of where things are heading. Cheap option generation makes it the constraint rather than removing it, and introduces a hazard — persuasive arguments for mediocre choices. Generate broadly and discriminate narrowly, write down the heuristics you can articulate, and protect the slow, consequence-facing work that produces the rest.