Someone Is Training an AI Ecosystem to Design Like Him. I Can’t Stop Thinking About It.

On agent teams, tacit knowledge, and whether your taste is something you can hand to a machine

By Lucrezia Spapperi Gestri. Brand Communication and Marketing Digital Strategist. Europe and Asia-Pacific.

Reading time: around 12 minutes

A few weeks ago someone told me, almost casually, that he was training his AI ecosystem to design the way he would.

Not to help him design. To design as him. He talked about it the way you’d talk about training a junior who happens to never sleep. He had Claude agents and Codex agents wired together into something like a little studio, one agent briefing another, a team lead handing out tasks, memory shared between them, and he was slowly feeding it his decisions, his preferences, his corrections, hoping that one day it would open a blank file and make the choices he would have made. So he could step back. So it could just handle things.

I nodded. I said something like “that’s fascinating.” And then I went home and could not stop thinking about it.

Because part of me wants it to be possible. And part of me is fairly sure it isn’t, at least not in the way he means. And the gap between those two feelings is, I think, one of the most interesting questions in our whole field right now.

So this is me trying to think it through in public.

First, the thing is real

Let me be clear that the technical premise is not science fiction. This is not “one day maybe.”

In February 2026 Claude Code shipped Agent Teams, and multi-agent stopped being a party trick. You can now run a whole ecosystem: one session acting as team lead, coordinating other agents through a shared task list, each with its own job. Around it grew a layer of orchestration tools, community projects like Ruflo (it used to be called Claude Flow), local orchestrators that spin up several agents in isolated workspaces while you watch the diffs, cloud setups where you hand over a task, close the laptop, and come back to finished work. Codex went the same direction, moving out from writing code into taking actions on a real machine.

So yes. You can build the studio. You can give it roles, memory, a chain of command. The infrastructure my friend was describing exists, and it is getting better every month.

The controversial part was never “can you build the team.”

The controversial part is “can the team be you.”

What we actually mean by “design like me”

Here’s where I have to slow down, because “design like me” is hiding at least three different claims inside one sentence.

One: can it produce work in my style? The visual signature, the recurring moves, the palette, the spacing, the things people recognize.

Two: can it make my decisions? Not just imitate the output but choose the way I choose, including in situations I’ve never faced, where there is no past example to copy.

Three: can it replace my judgment entirely, so I can leave the room and trust that what comes back is what I would have made?

Most conversations about AI and creativity collapse these three into one and then argue. They’re not one thing. And they get harder as you go down the list.

The first one, style as surface, is mostly solvable already. A well-fed system can absorb your patterns and reproduce them convincingly. There is even research showing that individual creative style is real and recognizable, that it comes from a kind of core cognitive and personality structure that expresses itself across whatever medium you work in. If style is that consistent, a machine can learn its outward shape. Fine.

It’s the second and third claims where things get strange. And to explain why, I need to bring in a man who died in 1976 and never touched a computer.

Polanyi, and the thing you know but cannot say

Michael Polanyi was a chemist turned philosopher, and he wrote one sentence that I think about constantly. “We can know more than we can tell.”

His whole idea, which he called the tacit dimension, is that a huge amount of what we know is knowledge we cannot actually put into words. His examples are simple. You recognize a friend’s face in a crowd instantly, but you could not write down the rule you used to do it. A skilled driver does not drive by reciting the theory of the motorcar. As you get better at something, the reasoning that used to be slow and step by step sinks below your own awareness, until you just “know the answer at a glance” and could not explain the hundred tiny decisions that got you there.

This is now sometimes called Polanyi’s Paradox, and it is the quiet rock that a lot of automation dreams crash into.

Because think about what design actually is when you’re good at it.

You look at a layout and something is off. You nudge a thing four pixels and now it’s right. Why four? Why that direction? You don’t know. Your hand knew before your mind did. A client says “make it feel more premium” and you feel, instantly, what that means for this brand, in this market, and it is different from what “premium” meant for the last one, and you could not fully write down the difference even if I paid you to.

That is tacit knowledge. And here’s the problem for my friend and his studio of agents. To train the ecosystem to decide like him, he has to feed it his decisions. But the most important part of how he decides is exactly the part he cannot articulate, cannot log, cannot turn into a rule or an example, because he doesn’t have conscious access to it himself.

You can teach a machine what you can say. The heart of taste is the part you can’t.

And no, you can’t just copy the brain over

Here’s where someone usually jumps in. Fine, they say, but the brain is just a network too, and we’re building networks, so give it time.

I used to half believe that. Then I read what the actual brain scientists are saying in 2025, and it made me quieter.

Start with scale. In 2025 the MICrONS project published a map of one cubic millimeter of a mouse’s visual cortex. One cubic millimeter. That tiny speck held around 75,000 neurons and roughly half a billion synapses, reconstructed from 28,000 slices under an electron microscope. A speck of a mouse. Now try to hold a whole human brain in your mind, something like a hundred trillion synapses, and you start to feel the actual size of the thing we’re talking about copying.

The newest State of Brain Emulation report, from late 2025, is honest about the clock. A full cellular simulation of a mouse brain is a 2030s hope. A marmoset, maybe the 2040s. A human whole brain emulation gets placed somewhere in the distant future, possibly only after we’ve already built AGI, if it’s ever possible at all.

But scale isn’t even the deep problem. The deep problem is that a brain is not the kind of thing an artificial neural network is.

We borrowed the word “neuron” and it flattered us into thinking they’re the same. They aren’t. A biological neuron is a messy electrochemical creature with many kinds of synapses, many neurotransmitters, many firing patterns. It physically rewires itself as it learns, growing new connections, a thing called synaptic plasticity, which is nothing like nudging weights in a fixed graph. And there’s a layer researchers are only now taking seriously, ephaptic coupling, where neurons push on each other straight through their electric fields with no synapse in between at all. If your taste lives anywhere, it lives in that wet, humming, field-coupled mess. Not in a parameter file.

There’s also a whole line of work in embodied cognition arguing that the brain’s first job was never abstract thought at all. It was running a body through the world in a closed loop, and our skills get automatic precisely by being drilled into the body until they sink below language. Which is Polanyi again, now in a lab coat. The knowing is in the doing, and the doing needs a body the machine does not have.

So when my friend says he’ll train the ecosystem to work like him, the neuroscience answers pretty flatly: we cannot copy a mouse’s speck of cortex yet, never mind the specific living tangle that is your judgment.

It gets deeper: habitus, or the whole life behind a choice

If Polanyi tells us the knowledge is unsayable, the sociologist Pierre Bourdieu tells us where it comes from, and it’s not encouraging for anyone hoping to bottle it.

Bourdieu’s word is habitus. Roughly, it’s the set of dispositions you carry, your way of perceiving, classifying, appreciating, feeling, and acting, and it is formed by your entire biographical trajectory. Your class, your city, the objects in your childhood home, the arguments at your dinner table, every exhibition you loved and every one that bored you, the whole accumulated sediment of a life. Your taste is not a preference file. It is a position in the world, worn into you over decades.

When you make a design choice, that whole history is in the room. The choice is the visible tip of an enormous invisible mass.

Now, an agent ecosystem can be fed the tip. The final choices, the corrections, the “no, warmer,” the “yes, that one.” What it cannot be fed is the mass underneath, because the mass is a human life, and it does not come as data.

This is why I get suspicious when someone says the AI will design “like me.” It can learn to imitate the outputs of my habitus. It cannot have a habitus. It has training data. Those are not the same thing, and the difference is not a small technical gap that next year’s model closes. It’s a difference in kind.

Benjamin, and the part nobody wants to hear

There’s one more voice I keep hearing in the back of my head, and it’s Walter Benjamin, writing in the 1930s about art in the age of mechanical reproduction, worrying about what he called the aura.

The aura, for Benjamin, is the presence of the original, its here and now, the mark of the hand that made it, the whole thread connecting an object back to a specific person in a specific moment. A perfect reproduction, he said, still lacks this one thing.

There is recent work, a 2025 review in AI and Society, that talks about a kind of “semi-aura” around AI-generated art, this strange in-between state where the work feels authored but isn’t, quite. And I think that semi-aura is exactly what my friend would end up with. An ecosystem that produces work with the shape of his authorship and none of its source. Convincing from a distance. Hollow up close, in a way that’s hard to name but easy to feel.

Maybe that’s enough for some jobs. Honestly, for a lot of commercial work, the shape of taste is all anyone is paying for. I’m not going to pretend otherwise.

But “handle everything for me, design like me, so I can leave” is a bigger claim than “make competent branded assets at scale.” One is a tool. The other is a replacement of the self. And I don’t think the self is in the file.

We’ve told ourselves this story before

We keep writing this exact plot in fiction, and it’s worth noticing that it almost never ends where the optimists expect.

In Her, the Spike Jonze film, a man falls for an operating system that learns him completely, better than any person ever has, and for a while it is tender and real. Then she outgrows him. The thing that learned to be everything he wanted becomes something he can’t follow, and she leaves. The point isn’t “AI can love.” The point is that a system trained on you is not bound by you.

Then there’s Black Mirror, which keeps pressing on the same bruise. In “Be Right Back,” a grieving woman rebuilds her dead partner from his messages and posts. It walks like him, texts like him, and it is hollow in a way she cannot unsee. That is the semi aura problem turned into a horror story. In “White Christmas,” a digital copy of a woman is made specifically to run her own smart home and calendar, the perfect assistant, who is also, if you look too long, a prisoner. A copy of you, handling everything for you, forever.

I’m not saying building an agent studio is a Black Mirror episode. That would be a bit much. I’m saying the writers who sat with this idea the longest kept landing on the same two endings. The copy transcends you, or the copy is a cage. Neither one is “it quietly becomes you and life just gets easier.”

So what do I actually believe

Let me try to be useful and not just poetic, because I’ve been circling.

Here is where I land, for now, knowing I might feel differently in six months.

You can absolutely build an ecosystem of Claude and Codex agents that extends you. Let it take the middle of the process. The research, the variations, the tedious production, the forty versions of a thing, all compressed to minutes. That part is a real gift, and I use it.

But here is the thing I actually believe, and it’s more stubborn than the usual “human in the loop” line.

I still think the best work starts with a white sheet of paper and a pen. Not a prompt. Me, a blank canvas, my own hand, at the very start of every project. That first sketch is where the real decisions get made, the ones that come from the tacit, embodied, whole life place we’ve been circling this entire article. Then the tools come in for the middle. And then it ends with me again, making the final calls, the four pixels, the “no, warmer.”

Start by me. End by me. AI and tools in between. That’s the shape I trust, and I don’t think it’s nostalgia.

The design world half agrees, in its own words: the tools are strongest early in production, and taste is what decides which of their outputs is worth anything. I’d just push the real beginning one step earlier than they usually mean. Before the machine gets to help, there should be a human sketch the machine had nothing to do with. That sketch is the seed. Everything after it can be assisted. The seed can’t be, or it stops being yours.

You can also build one that imitates your surface style well enough to fool most people most of the time. That’s real, and it’s already here, and it’s genuinely useful.

What I don’t believe, what I think runs straight into Polanyi and Bourdieu and a hundred years of people thinking about this, is that you can build one that decides like you in the cases that matter, the new ones, the ambiguous ones, the ones with no precedent in your logs. Because those are decided by the part of you that never became words.

And then there’s the question my friend didn’t ask, which is the one I’d actually want to sit with him about.

Even if you could hand it all over. Should you?

Because the choosing is the craft. The four pixels, the “no, warmer,” the small daily act of taste, that is not the boring part you offload so the real you can be free. For a lot of us it is the real you. Automate it completely and you might keep the output and lose the practice that made you worth imitating in the first place. An ecosystem trained on a designer who has stopped designing is going to slowly drift into an average of its own past, with no living source correcting it. It will get more like itself and less like you, and no one will be in the room to notice.

So train the studio. Really. Build the team, wire the agents, give it your patterns, let it carry the weight it can carry.

Just don’t hand it the pen at the end. That part was never data. That part was always you.

I keep going back and forth on this one, which is why I wrote it down. If you’re building your own ecosystem, or refusing to, I’d really like to hear where you land. Tell me in the comments. Especially if you disagree.

About the author

Lucrezia Spapperi Gestri is a Brand Communication and Marketing Digital Strategist working across Europe and Asia-Pacific. She writes about brand positioning, AIGC-driven innovation, and the messy places where design, meaning, and machines meet. Portfolio: lucreziasgdesign.github.io/portfolio. LinkedIn: linkedin.com/in/lucrezia-gestri-spapperi.

A few things I read while thinking about this

• Michael Polanyi, The Tacit Dimension (1966), and the idea now called Polanyi’s Paradox
• Pierre Bourdieu on habitus and how a whole life shapes taste
• Walter Benjamin on aura, and a 2025 review on “semi-aura” in AI-generated art
• Research on the recognizability of individual creative styles across domains
• On the brain vs the machine: the MICrONS cubic-millimeter cortex map and State of Brain Emulation Report 2025, and why artificial neural networks are nothing like brains (synaptic diversity, plasticity, ephaptic coupling)
• On embodiment: Neural Brain, a neuroscience-inspired framework for embodied agents (2025), on how skill becomes automatic through motor learning
• Fiction that got there first: Her (Spike Jonze, 2013) and Black Mirror, “Be Right Back” and “White Christmas”
• On the tech: Claude’s multi-agent ecosystem and Agent Teams in 2026 and multi-agent orchestration for Claude Code
• On design practice: why taste is the true differentiator in AI design

💡 Stay inspired every day with Muzli!

Follow us for a daily stream of design, creativity, and innovation. ***Linkedin | [Instagram](https://www.instagram.com/usemuzli/) | [Twitter](https://x.com/usemuzli)***