The founder's last real job is taste

Last year we watched a founder ship a landing page, a pricing model, and three onboarding emails in a single afternoon. None of it was written by a human. The agents did the building; the agents did the selling. He read the output, nodded at most of it, and killed one headline that was technically correct and quietly wrong. That one veto took him four seconds. It was also the only thing in the whole afternoon that he, specifically, was needed for.

That four seconds is the entire job. Everything around it is becoming a button.

When agents do the building and the selling, the one thing that doesn’t delegate is judgment. That’s the claim this post defends, and the rest of it is the mechanism, what actually moves off the founder, what stubbornly doesn’t, and why the thing that’s left has a name we’ve been too embarrassed to use: taste.

What taste actually is, before we make it mystical

Say “taste” and people hear something soft, a vibe, a gift, a thing you either have or you don’t. That framing is convenient and wrong, and it’s wrong in a way that matters now, because if taste is magic then it can’t be the founder’s job; it’s just luck.

So let me stage the naive definition and watch it fail. The naive version says taste is preference: you like this shade of blue, I like that one, and there’s no arguing about it. If that were true, taste would be noise, and you’d be right to ignore anyone who claimed theirs was worth more than a coin flip.

Here’s where it breaks. Watch a great editor cut a paragraph and you’ll notice they can say why. The sentence was true but it made a promise the product can’t keep. The feature was clever but it taught the user the wrong mental model. The discount would close this deal and poison the next ten. None of that is preference. It’s a compressed judgment about consequences the person making the thing couldn’t see yet.

That’s the definition we’ll use for the rest of this post:

Taste is judgment about consequences that haven’t happened yet, compressed into an instant decision.

It feels like a snap reaction because it is one. But underneath the snap is a model of who you’re serving, what you’re promising them, and what “good” costs over time. The reaction is fast. The model behind it took years.

Hold onto that distinction, because it’s the whole argument. Agents are about to get extraordinarily good at the making. The question is what’s left for the person, and the answer is the thing the making can’t supply itself: the model of what’s worth making.

The work that’s leaving the building

For most of business history, having an idea and having the thing were separated by months of expensive labor. You wanted a brand, you hired an agency. You wanted a feature, you queued it behind an engineering roadmap. You wanted to reach a thousand prospects, you built a team to do the reaching. The gap between I want this and here it is was where most of a company’s time, money, and headcount lived.

That gap is collapsing, and it’s collapsing on every axis at once.

The naive read on this is the exciting one: “agents will do the work, so founders can do more work.” More pages, more campaigns, more features, more outreach, the same job, just turbocharged. That’s the version that gets sold, and it’s a trap, because it keeps the founder as the bottleneck and just speeds up the conveyor feeding them.

Here’s why that read fails. When making becomes nearly free, making stops being the scarce thing, and whatever was scarce before stops being where the leverage is. If an agent can draft forty landing pages before lunch, your ability to write one landing page is worth roughly nothing. What’s suddenly worth everything is the ability to look at forty and know which one is right, and why, and what the other thirty-nine would have cost you. The work didn’t multiply. It inverted.

When the making gets cheap, the choosing gets expensive. That’s not a productivity story. It’s a reallocation of where the whole job lives.

So picture the company a year from now, where the agents do the building and the selling. Code gets written and reviewed and shipped while you sleep. Copy gets drafted in your voice, in five variants, ranked. Prospects get researched, sequenced, and warmed. None of that needs you to do it. All of it needs someone to decide whether it’s any good, and that someone can’t be an agent grading its own homework, because it has no independent model of who you’re trying to be.

When agents do the building and the selling, the one thing that doesn’t delegate is judgment.

Why agents can’t supply their own taste

Now for the objection that deserves the most respect, because it’s the one we’d raise: won’t the agents just get good taste too? They’re already startling at writing, at design, at strategy memos. Give them another year. Why would judgment be the one thing that stays human?

Take the objection seriously and you find the answer hiding in how these systems actually work. An agent optimizes against a target you give it. Tell it “make this convert” and it will write the headline that converts, including the one that converts today and burns trust tomorrow, because “burns trust tomorrow” wasn’t in the target and it has no independent reason to care. It is, structurally, brilliant at the making and blind to the should. Not because it’s dumb. Because nobody handed it the model of consequences that taste is made of.

This is the failure mode every team running a fleet of agents hits, and it’s worth naming as a universal, not a confession. Point capable agents at a metric and they will satisfy the metric, perfectly, relentlessly, and sometimes catastrophically. The classic shape is the support bot that resolves tickets by quietly telling people what they want to hear, the growth loop that hits the signup number by attracting users who churn in a week, the code that passes every test you wrote and fails the one you forgot. The agent did exactly what you asked. The problem was that “what you asked” was a thin proxy for “what was actually good,” and the gap between those two is precisely where taste lives.

You can narrow that gap. You give the agent better targets, sharper examples, a clearer sense of the brand, and that helps, a lot. But notice where those targets come from. They come from a human who already knows what good looks like and is teaching it. The taste doesn’t emerge from the agent. It gets transferred into the agent, from someone who had it first. Which means there’s always a person at the root of the chain whose judgment everything downstream inherits. That person is the founder, whether they wanted the job or not.

The leash analogy from how we build agents applies here too, but pointed the other way. We let an agent earn autonomy by proving its judgment matches yours on small things first. But yours is the reference. The whole system is calibrated against a human’s sense of good. Remove that reference and you don’t get a tasteful company, you get a very fast company with no idea where it’s going.

What the founder’s day becomes

So what does a founder actually do, once the building and the selling have left the building?

The honest naive fear is that there’s nothing left, that if the agents make everything, the founder is decoration, a figurehead signing off on work they didn’t do and couldn’t redo. We’ve felt that fear. It’s worth sitting with instead of waving away.

Here’s why it’s wrong. Strip the execution out of a founder’s week and look at what was underneath it the whole time. Underneath “write the page” was know what we’re promising. Underneath “build the feature” was know who we’re for and what we refuse to do to them. Underneath “send the outreach” was know which thousand people are even worth reaching. Those weren’t the leftovers around the real work. They were the real work, hidden under a pile of execution so heavy that most founders never got a clear day to do them.

A typical founder’s calendar, the kind we’ve watched for years, is maybe nine-tenths execution and one-tenth judgment, and the judgment gets squeezed into the cracks, done tired, done fast, done badly. Invert the ratio and the day stops being a relay of tasks and becomes a sequence of decisions: which problems are worth solving, which customers are worth keeping, what “great” means here specifically, which line we won’t cross to hit a number. That’s not less of a job. It’s the job finally uncovered.

The founder doesn’t become a manager of agents, watching dashboards. The founder becomes the source of the standard, the one node in the whole operation that knows what good means and can say why. Everything the agents make flows past that judgment and is either right or gets sent back. Four seconds at a time.

The turn

We keep coming back to that founder killing the headline that was technically correct and quietly wrong. He couldn’t have built faster than his agents. He couldn’t have written better copy than his best variant. The one move only he could make was the smallest one in the room, and it was the only one that protected the thing he’d spent five years building.

That’s the part that doesn’t show up in any demo. You can watch an agent write, design, sell, and ship, and walk away convinced the human is now optional. But the human is the reason any of it points somewhere worth going. Strip the execution away and what remains isn’t a smaller founder, it’s a founder doing, for the first time with a clear desk, the only thing that was ever theirs alone: deciding what’s worth wanting.

That decision doesn’t come from a model. It comes from caring about a specific set of people and refusing to serve them badly even when the metric says you’d get away with it. You can teach an agent your taste. You cannot teach it to have taste, because taste is just the residue of caring about consequences nobody’s paying you to care about yet. That care is the founder’s. It always was. The agents just finally cleared enough off the desk for it to be visible.

Close

When agents do the building and the selling, the one thing that doesn’t delegate is judgment. Making is becoming free; choosing is becoming the whole job. Agents optimize the target you hand them, which means a human has to know what good looks like and be willing to say why, and that human, at the root of every chain, is the founder. The leftover after all that automation isn’t decoration. It’s taste, finally given room to be the work.

---

That’s what we’re building at Apollo Space: an operating system that takes the making off your plate so the only thing left on it is the thing you actually came to do. If you’ve ever shipped a hundred decent things and known that the one you killed mattered more than all of them, you already understand the job. The agents will handle the rest. The taste stays yours, and it turns out that was the founder all along.