Taste is a feature

There is a simple test for any work an agent hands back: would you forward it without editing? If the answer is yes, the agent saved you time. If the answer is no, it just moved the work around. You still have to rewrite the thing before anyone else sees it.

That test is why we treat taste as a feature, not a coat of paint applied at the end. For an agent, taste is the line between output you send and output you fix.

Thoughtful over eager

An eager agent answers the question you asked and stops. A thoughtful one notices the thing you did not ask about: the date that does not line up, the claim that needs a caveat, the better way to phrase the ask. It does not over-explain to prove it was working. It says the useful part and trusts you to take it from there.

We tuned for that on purpose. Two of the three evals we hold the product to are about exactly this: the quality and look of what it produces, and how intelligent it feels to work with. Noticing the non-obvious counts. So does knowing when to be brief.

Reads like a person

Most of taste is restraint. Dyeu writes in clean Slack formatting, with the bold and the lists where they earn their place and not a sentence more. No throat-clearing, no padding, no chatbot tics that announce a machine wrote this. It matches your voice because it has read your company, not because it is performing friendliness.

The bar

The bar we hold ourselves to is the forward test. Could you take what the agent wrote, drop it in front of a customer or your team, and feel good about it. Not "good enough to fix," but good enough to send.

Hitting that bar is not polish you add at the end. It is judgment about what to say, restraint about what to leave out, and care about how it reads. We treat all three as core, because for an agent, that is the product.