Operating model · Agents

What a one-person creative team looks like when the team is agents

By Orcool · July 20, 2026 · 5 min read

The team got small. The work did not.

A few years ago, a serious creative operation needed a room. A researcher to watch the category, a strategist to shape angles, writers and editors to produce, someone to sort the pile and decide what actually ships. Each of those was a person, and coordinating them was its own job.

Most of those roles have quietly become software you can talk to. You can hand a category to an agent and get back what competitors are running. You can hand it real user reviews and get back angles written in the words customers actually use. You can ask a second agent to rank those angles before a cent of media sits behind them. What used to be a room is now one person in a chat window with a small crew of agents on the other side.

The interesting part is not that the team shrank. It is what stayed, and where the human ended up sitting.

Three seats the agents fill

It helps to name the roles, because a team of agents is still a team. It has a shape.

The researcher

The first agent reads the market and gives it back as structured signal: which hooks competitors are running, which formats are holding up in the category, what real reviews say in real language. This is the least glamorous seat and the one that decides everything downstream, because an angle is only as good as what it was built from. We wrote more about this shift in the creative producer becoming a GTM engineer: the job moved from briefing editors to briefing signal.

The drafter

The second agent turns that signal into angles. Not a blank-prompt script that could belong to any app, but a ranked set of ideas that each trace back to something real: a frustration from a review, a hook the category is already rewarding, a format with room left in it. The draft is the output of the research, not a fresh guess dressed up as one.

The critic

The third agent scores the angles against a creative-effectiveness framework before anything goes live, so the human opens a shortlist instead of a slush pile. In one blind five-week test we published, an angle built this way scored 0.76 on Google's ABCD framework against 0.49 for a single-tool draft of the same app. The point is not the exact figure, it is the shape: a critic seat lets you rank before you spend rather than after you have burned budget finding out.

researcher reads the market  →  drafter turns signal into angles  →  critic ranks before spend  →  you decide

The one seat that stays human

Notice which chair is still empty. The agents research, draft and rank. None of them decides. The yes or no on an angle, which one is right for this brand, this moment, this audience, is the seat that does not hand off. And it is not a bottleneck to apologize for. It is the whole reason the setup works.

Taste is the part that does not scale, because it is specific. A generic model can tell you what tends to win across everyone. It cannot tell you what your brand would be caught dead saying, or the joke that lands with your exact audience and nobody else's. That judgment lives in one head, and the best a system can do is learn it. When your decision teaches the pipeline what to bring you next, the research gets sharper and the shortlist gets shorter. That loop is what we call the Brand Lens: the decision-maker's yes and no, captured and reused.

The one-person team does not work because the human does less thinking. It works because the human does only the thinking no agent can do for them.

Why more agents do not remove the bottleneck

There is a tempting version of this story where you add enough agents and the human falls away entirely. That is not what happens, and it is worth being honest about why.

Every agent you add produces more to judge. A drafter that returns twice as many angles hands you twice as much to say yes or no to. If judgment is the constraint, more production pushes harder against it, not past it. The move that actually helps is the opposite: organize the agents around the decision, so the research feeds the drafts, the drafts get ranked, and the human only ever touches the short, ranked end of the funnel. The agents do not replace the seat. They defend it, by making sure the only thing that reaches it is worth a decision.

This is also why the honest measure of a one-person team is not how much it produced. Output was never the scarce thing. The scarce thing is a winning angle that holds, and the number worth watching is how fast you get to one and how long it lasts. We call that reaching the winner in the tool you already have open rather than in yet another dashboard, because a one-person team cannot afford a context switch for every step.

What this looks like on a Tuesday

Concretely: one person opens a chat. They ask the researcher what shifted in the category this week and get three competitor hooks and a fatigue read back. They point the drafter at a spike in reviews mentioning the same small frustration, and get five angles that lead with it. The critic ranks them and flags two as weak at the hook. The human reads the top three, kills one for being off-brand in a way no score would catch, greenlights one, and sends it to build. Fifteen minutes, one person, a full creative cycle, and the only irreplaceable moment in it was the kill and the greenlight.

That is the quiet promise of a team of agents. Not that you stop making decisions, but that decisions become almost the only thing you do.

Frequently asked questions

Methodology note: the 0.76 vs 0.49 figures come from a single internal head-to-head run over five weeks on one app, scored blind on Google's ABCD framework, first published in our ABCD benchmark post. The client is anonymized and the competing tool is described by category. A single test is a signal, not a law, and the number is used here to illustrate the critic seat, not as a general claim.