The first thing I saw on the whiteboard at Maesa was an org chart. Not their org chart — an agent org chart. Someone had drawn boxes labeled "creative director agent," "copywriter agent," "designer agent," "producer agent," with dotted lines between them, and it looked exactly like the team page on their website. They were proud of it. They had spent two weeks on it. It had produced almost nothing anyone could use.
So we tore it up and did something duller. We walked a single brand launch backward, step by step, and wrote down every place work actually stopped and waited. Not roles. Steps. Naming shortlist to legal screen. Concept to packaging copy. Packaging copy to retailer spec check. Spec check to line review. There were somewhere near thirty of them, and about six were where all the time went. Those six got agents. The other twenty-four did not.
That's the version that worked. Maesa now runs launches at roughly $280K less per launch, in 3 months instead of 9, across 12+ brands. Oshyia Savur, their VP of Marketing, put the number at "tens of millions" from the stage at Shoptalk last year. None of that came from a creative director agent. It came from a handful of narrow, unglamorous specialists sitting at the exact points where the line used to jam.
Every named the pattern. Most teams build it backward.
Every's Context Window ran a piece on August 2 called "Your AI Is a Team of Specialists", arguing that these systems work better as a set of narrow roles than as one model doing everything. I think that's right, and I'd add the part that decides whether it works in a creative org: which specialists you pick.
Because there are two ways to decompose a team, and only one of them is real.
The org chart is a coordination artifact. It exists so people know who approves what and who to escalate to at 6pm on a Friday. It is a map of accountability, not a map of work. When you copy it into agents, you inherit its abstractions without inheriting the thing that made it function — the humans who quietly route around it every day.
The production line is the actual work. It's the sequence of artifacts that get made, checked, revised, and handed off. It's specific, it's boring, and it's where every real delay lives. Agents built against it have something a "creative director agent" never has: a defined input, a defined output, and a way to tell whether the output was any good.
The tell: you can't say what "correct" looks like
Ask someone to define success for a copywriter agent. You'll get adjectives. On-brand. Sharp. Fits the voice.
Now ask them to define success for the thing that checks packaging copy against a retailer's spec sheet. You'll get a list. Character counts per field. Required claims present. Banned claims absent. Ingredient order matching the regulatory format. Territory-specific variants generated.
The second one is buildable. You can write the eval, run it against forty past launches, and know before you ship it whether it's better than the person who used to do it at 11pm. The first one is a vibe with a job title attached.
That's the practical difference. Specialists defined by production step come with their own scoreboard. Specialists defined by job title come with an argument.
Narrow beats smart
There's a reflex on creative teams to build the impressive thing first — the assistant that knows the whole brand and can do anything you ask. I understand the appeal. It demos beautifully. It also degrades the moment two people use it for two different jobs, because everything you add to make it better at one task makes it worse at another.
The narrow ones don't have that problem. Across the engagements we've run, the pattern holds: average production time down about 90%, and it almost never comes from one heroic system. It comes from six or eight small ones that each do a single, checkable job and hand off cleanly.
The other thing narrow buys you is ownership. When an agent maps to a step, the person who owns that step owns the agent. They wrote the checks. They notice when the output drifts. Compare that to a shared brand assistant that belongs to everyone, which means it belongs to no one, which means in six weeks it's quietly stale and people have gone back to doing it by hand while still telling you in the standup that they're using AI.
How to find your six
If you want to run this yourself, this is the whole exercise. It takes an afternoon.
Walk one real project backward. Not a hypothetical. A specific job that shipped last quarter, with the people who worked on it in the room. Start at delivery and go back to brief.
Write down every handoff and every wait. Every point where an artifact sat in someone's inbox or slack thread. Where it went back for a second pass. Where someone rebuilt something that already existed.
Mark the ones that recur. Once is a project. Every launch is a system. You only want the recurring ones.
For each candidate, write the correctness test before you write the prompt. If you can't state what a good output looks like specifically enough to check it, that step isn't ready to be an agent. Move on. Come back to it.
Build the top three. Not thirty. Ship them, watch them for two weeks, then take the next three.
What you'll notice, and what caught the Maesa team off guard, is how few of the winners look like creative work. The line jams at spec checks, asset resizing, version reconciliation, territory variants, and the endless translation of one approved thing into eleven required formats. That's where the months went. Nobody's org chart has a box for it, which is exactly why nobody built an agent for it.
Your team's structure tells you who's responsible. Your production line tells you where the work actually is. Build for the second one.