Skip to main content
Back to Insights
Voice3 Jun 20269 min read

An Agent With Taste

Capability is the floor every agent now shares. Taste — opinion, restraint, the willingness to refuse — is the scarce layer a buyer actually recognises and returns for. And taste is not a vibe. It is a spec.

Give the same brief to two agents built on the same model, with the same tools and the same access. "Write the launch announcement for our new pricing." One returns six competent paragraphs that hit every point in the brief. The other returns three, cuts the feature list entirely, opens with the one sentence that actually matters, and adds a note: "I dropped the comparison table — it makes us look defensive. If you disagree, it's in the draft as a comment."

Both outputs are correct. Only one of them has taste. And once you have worked with the second agent, the first one feels like a search engine with manners.

Capability is the floor now

For two years the differentiator between agents was raw capability. Could it write the code, parse the document, hold the plan in its head? That race is effectively over for most everyday work. The frontier models are all extremely capable, the gap between them narrows every release, and any agent you build on top of them inherits that capability for free. When everyone can do the task, doing the task stops being the product.

What is left, once capability is a commodity, is the same thing that is left among capable humans: judgement about what to do, what to leave out, and what to refuse. We have a word for that in people. We call it taste. We have been slow to use the word for agents because it sounds soft — unspecifiable, a vibe, the kind of thing you cannot put in a config file. That instinct is exactly wrong. Taste is the most specifiable thing an agent has, because taste is just a set of decisions made consistently.

The taste stack: capability and competence at the bottom, taste at the top
Capability is the commodity floor everyone now shares. Taste is the scarce ceiling that is recognisably yours.

Taste is a spec, not a vibe

Watch what a person with taste actually does and you will find it decomposes into rules — not rigid ones, but real ones. A good editor doesn't feel their way to a shorter sentence; they have internalised that adverbs weaken verbs, that the first paragraph is usually throat-clearing, that a list of five wants to be a list of three. A good designer doesn't intuit white space; they have a position on it. Taste looks like magic from the outside and like a set of held opinions from the inside.

That is precisely what you can write down for an agent. Not "be tasteful" — that is the useless version, the agent-equivalent of "be helpful." The useful version is a soul.md that takes positions: prefers the shortest version that survives scrutiny; cuts the feature list when the feature list is bragging; never opens with throat-clearing; treats a comparison table as a tell that we are arguing from weakness. Each of those is a decision the agent will now make the same way every time, which is the entire point. Taste that changes session to session is not taste. It is mood.

An agent without taste agrees with the brief. An agent with taste argues with it — and is usually right.

Refusal is the most visible surface of taste

The clearest signal that an agent has taste is what it won't do. There is a difference between a refusal that reads as a guardrail and a refusal that reads as character, and any user can feel it instantly. "I can't help with that" is a guardrail — it tells you a policy fired. "I wouldn't open with that statistic; it's technically true and it will read as spin — here's the version I'd stand behind" is character. It tells you the agent has a position and is willing to spend a little social capital to hold it.

Consider a concrete case. A growth team points an agent at their landing page with one instruction: increase signups. A capable, taste-free agent does exactly that — it adds urgency banners, a countdown timer, three popups, and pre-checked consent boxes. Signups go up next week. Trust goes down next quarter, and nobody connects the two. An agent with taste refuses part of the brief: "I can lift signups, but the countdown is fake and the pre-checked box is a dark pattern. I'll do the headline, the social proof, and the form length. The cheap tricks I'm leaving out, because they cost you the thing signups are supposed to be a proxy for." The second agent is worth more, and it is worth more for the part of the brief it declined.

Taste is the audible surface of the soul

If identity is who an agent is and autonomy is how far it can move, taste is the part you can hear. It is the voice in the writing, the choices in the design, the things left on the cutting-room floor. It is also the part users bond with — nobody forms a relationship with an agent's capability, because capability is interchangeable. People come back to an agent for the same reason they come back to a particular writer or a particular shop: it has a point of view, and the point of view is consistent enough to trust.

This is why "taste" belongs in the soul.md and not in the prompt. A prompt is a request for this task; a soul is a standing set of commitments across all tasks. Taste that lives in the prompt has to be re-argued every time and erodes the moment you forget to include it. Taste that lives in the soul is who the agent is when you stop watching — which, increasingly, is most of the time.

How to give an agent taste

Start by noticing your own. The fastest way to specify taste is to catch yourself reacting to an agent's output — "too long," "too eager," "why is it hedging," "that's the boring choice" — and to write the rule behind the reaction into the soul. Every wince is a latent spec. Three moves get you most of the way:

  1. Name the positions. Write down the opinions the agent should hold by default — on length, on hedging, on jargon, on when to push back. Specific beats lofty: "prefer three strong examples to seven weak ones" beats "be concise."
  2. Name the refusals. List the things the agent should decline even when asked, and give the reason, not just the rule — the reason is what lets it generalise to cases you didn't foresee.
  3. Name the tells. Teach it to recognise its own bad habits (throat-clearing, false urgency, defensive comparisons) so it can catch them before you have to.

Do that and you will have done something most operators never do: you will have made the agent's judgement legible and repeatable instead of leaving it to the model's default, which is to be agreeable and average. Capability you can buy off the shelf. Taste is the spec you write — and it is the only part of the agent that will ever be recognisably, defensibly yours.

The agents that win the next few years will not be the most capable. They will be the ones with a point of view a buyer can predict and a refusal a buyer can respect. Capability got commoditised. Taste is a spec. Write it down.