Dieter Rams, one of the greatest designers no one other than designers has ever heard of, spent four decades at Braun making mundane things people used without thinking about them — yet the radios, calculators, shavers, the shelving Vitsœ still sell.

Elegant in their simplicity and clarity.

For Rams, beauty wasn’t applied to function — it emerged from it. The work was understanding people deeply enough that the right form became obvious.

In the 1970s he boiled his approach down to ten principles for good design. Much of the design field, including myself, has run on them quietly ever since; most people don’t understand how much Rams has influenced the devices they use. His minimalist-functionalist canon is a winning strategy, not the winning strategy.

His won still. A lot.

The point: He didn’t invent functionalism; he framed it through the restraint of ten principles so it was clear and easy to understand.

That framing is massive, with Apple’s product line and $4.5 trillion market cap as an obvious example.

One of his biggest disciples, Jony Ive, built much of Apple’s design language on top of them, stealing without apology with Steve Jobs as an advocate. You’ve probably recited a couple in a critique without knowing whose they were.

I’ve held them up repeatedly, and they’ve always been where I go to for designing software and quite frankly, living my life.

Then there’s the economic impact.

If you assigned even 5–10% of Apple’s brand premium to design-language inheritance, you’d be at a few hundred billion dollars. From one company.

Then it fans out.

• Muji is basically Rams in retail form.
• Vitsoe is his actual furniture, still selling shelving 60 years later.
• Design Within Reach, one of my favorite stores.
• Braun itself. At a premium.

Then every minimalist consumer hardware brand of the last 20 years — Nest, Sonos, Dyson (different in form but same functionalist DNA), every smart home device that’s a white rounded rectangle.

On the software side, the entire flat-design turn in the 2010s — iOS 7 onward, Material Design’s more restrained moments, the whole “less but better” tic in SaaS UI — pulls from the same well.

Decades of influence. One man.

Here’s the part that should give us pause: Rams, in his nineties, doesn’t own a computer and probably never has many of the devices he influenced.

My guess he didn’t care.

That’s because he understood it wasn’t about the technology, was about the human outcome. The user doesn’t care whether a feature runs on a specific model or not; they care whether it works, whether it helps, whether they can trust it to accomplish the task they need done.

Because of this, I believe good AI design is just good design — which is why a 1970s list still works because it is long-lasting. That’s principle 7.

Good Design Is Innovative

Rams tied innovation to new technology, but he never treated novelty as the for the lede of the story. A new capability was a reason to rethink the product that made it easier to use, not a license to decorate it.

Real innovation asks what the technology lets you remove, collapse, or rethink about the underlying job. The honest version of that is uncomfortable, because it often means the visible feature gets smaller, not larger, or disppears altogether like an ambient agent. A genuinely new use of a model might collapse a five-step form into one sentence a user types, or delete screens entirely — outcomes that don’t photograph well in a launch deck.

And this alone is why many designers are struggling in this new age — it’s about reducing complexity, not adding it. That’s hard.

Novelty is loud. Innovation should be quiet, and yet most organizations reward loud. Don’t be them.

Often the best use is taking a mundane task the user quietly dreads — reconciling records, cleaning a messy list, reformatting a document — and making it disappear or making it simple.

Rams’ radios weren’t innovative because they used the newest transistor; they were innovative because the new component let him simplify the object around it. The same discipline applies now. Ask what the model makes newly possible to take away, not only what it makes possible to add.

Solving the boring well is innovation, and usually a more useful one than the demo that gets the applause.

Designing for AI

• Map the workflow first. The model’s value is in changing the shape of the work, not sitting beside it — and you can’t see that opportunity without understanding the job first.
• Ruthlessly cut. Keep reducing the steps, and ask if AI can help do that, like the five whys.
• Test the baseline. A model that can’t beat a plain rule or default is just cost and unpredictability with nothing to show for it.

Good Design Makes A Product Useful

Rams said design exists to make a product useful, and to throw out anything that gets in the way of that.

The AI-era reflex runs backward: you add a capability because it kills in the demo, then go looking for someone who needs it. Usefulness is narrower and harder than that. It means the feature shortens the gap between what a person wants and getting it, measured in their work.

Usefulness is unforgiving in a way that raw capability isn’t. A demo only has to work once, on the happy path you chose. A useful feature has to work on the inputs people actually bring, on the days they’re rushed and imprecise.

Most AI features clear the first bar and quietly fail the second, and the gap between “it worked in the demo” and “it works in the job” is where adoption dies.

A demo has an audience and a date. Usefulness shows up weeks later, in someone else’s workflow, with no one in the room to defend it — which is how it loses, quietly, to the thing that looked good in the review.

There’s a subtler failure underneath that one: features useful to the company but not the user. Engagement climbs, the model gets more sessions, and none of it shortens anyone’s actual task. Rams’ test was simpler and harder — does this help the person do the thing they came to do? Hold AI to that, and a lot of shiny features don’t make the cut.

Designing for AI

• Define the task. A feature with no named job it shortens is a capability hunting for a use, and engagement numbers will flatter it long after it has stopped helping anyone.
• Test on the mess. The demo runs on the inputs you chose; the product runs on the ones people actually bring, rushed and imprecise, and that gap is where usefulness is won or lost.
• Build the escape hatch. A probabilistic system will be wrong sometimes; if fixing its mistakes costs more than doing the task, the feature is a net loss however good it looks.

Good Design Is Aesthetic

Rams treated aesthetics and function as the same conversation; a well-made thing is pleasant to live with, and that pleasantness is part of how it works.

In AI products, the aesthetic question has moved the answer itself.

Most generative interfaces converge on the same blank chat box — the ChatGPT text field that nearly every product now clones — which is the aesthetic equivalent of shipping everything in gray.

When every product answers with the same field and the same streaming cursor, the interface stops telling you anything about what it’s good at or how to use it well. A calculator that looks like a chat thread teaches you nothing about calculation; a coding tool that shows its edits as a diff you can accept or reject teaches you everything you need to know about how it works.

It’s more about the interaction — how fast a response comes back, the tone of the copy, what your screen does to build trust while the model is thinking.

Pacing and content restraint now matters as much as layout. A model that dumps a wall of text the instant you stop typing feels different from one that pauses, structures its answer, and shows you where it’s going — even when the underlying output is identical. Words matter more than icons.

Rams understood that how a thing behaves is part of how it looks. In AI, behavior is almost all of the aesthetic.

Designing for AI

• Meet the user where they are at. The output surface tells users what the tool is for and the intent of the interaction. Slack needs a much different approach than the web, so design with that intent.
• Design the stream. How an answer arrives — its pace, the way it takes shape — shapes how much it’s understood and trusted, not just the words that land.
• Fill the empty state. A blank field hands the user the job of guessing what the tool can even do, which is exactly the moment most of them give up. Set expectations with intent.

Good Design Makes A Product Understandable

Good design, Rams wrote, explains itself. AI fights that instinct at every turn. The outputs are probabilistic, the reasoning is a black box, and your users are left staring at an answer wondering where it came from. Understandable now means making the system explainable: what it can do, where it stops, and why this particular answer showed up.

A large part of that is setting expectations before the user ever sees an answer. An understandable system signals what it’s good at and what it isn’t, so people arrive with calibrated expectations instead of discovering the limits the hard way.

It’s scaffolding with purpose.

Many AI products do the reverse — an open prompt and a confident tone imply the model can handle anything, and the disappointment shows up later, quietly, at the boundary. Scope stated up front sets an expectation the rest of the experience can actually meet.

The hard part is that the system itself often can’t explain why it produced an answer, so understandability has to be designed around the model rather than extracted from it.

That means surfacing what it drew on, what it’s confident about, and what it’s guessing — scaffolding the model won’t volunteer on its own.

The interface has to do the explaining the model can’t.

Perplexity does a version of this: it anchors each answer to numbered sources you can open, so the basis for a claim is one click away rather than something you take on faith. The provenance is part of the surface, not bolted on afterward.

Most products skip this because the fluent answer feels self-evident. It isn’t.

A user who can’t tell where an answer came from also can’t tell when to trust it, and that uncertainty erodes the whole product slowly. A confident paragraph with no visible basis is the least understandable interface we’ve shipped in years, and legibility is what lets people use the system at all.

Designing for AI

• Set expectations up front. Expectations form whether you manage them or not; an open prompt and a confident tone quietly promise everything, so the limits, met later, land as a betrayal.
• Expose the steps. An agent that acts invisibly is asking for trust it hasn’t earned — people can only rely on what they’re able to check.
• Explain the answer. With no visible basis, a user can’t tell a sound answer from a confident guess, so they either over-trust the system or abandon it — and both are failures.

Good Design Is Unobtrusive

Rams compared good products to tools: neutral, restrained, leaving room for the person using them. AI pulls hard the other way. It wants to suggest, autocomplete, summarize, and act, usually before anyone asked. Unobtrusive AI waits. It sits lambently in the background until the moment it’s actually useful, and it doesn’t hijack the intent the person walked in with.

The pressure runs the other way, though. Every team wants to show that its AI is “doing something,” so the assistant pops up, pre-fills, and narrates — proving its presence at the cost of the user’s attention.

Activity gets mistaken for value.

A tool that constantly reminds you it’s there is, by Rams’ standard, poorly designed.

Restraint is hard to ship because it’s invisible on a roadmap. “We made the assistant interrupt less” is not a headline feature, even when it’s the right call. But the products people keep using tend to be the ones that stay quiet until needed.

GitHub Copilot is a decent model of this: it offers code as dimmed ghost text right at the cursor, which you take with a keystroke or erase simply by continuing to type — there when wanted, gone when not. The difference between an assistant and a nuisance is mostly a question of when it speaks.

Designing for AI

• Surface where it matters. Help the user didn’t ask for competes with the work they’re trying to do; attention is the scarce budget every interruption spends.
• Keep suggestions light. Anything the model writes into a user’s work uninvited turns them into an editor of your output instead of the author of their own.
• Offer an off switch. A proactive feature you can’t turn off isn’t a tool, it’s an imposition — and the freedom to refuse it is what keeps the user in charge.

Good Design Is Honest

This is the one AI tests hardest. Rams meant design shouldn’t make a product look more capable or valuable than it actually is. Chat interfaces do exactly that — a fluent, confident tone reads as competence even when the answer is flat wrong.

When Columbia’s Tow Center tested eight AI search tools on naming real news sources, they were wrong more than 60% of the time, and the tell is that they rarely just declined — they guessed, confidently. That’s not honest.

We can put the signal back and explainability is the main instrument for doing it. Showing the basis for an answer — the sources, the steps, the data it leaned on — lets a user weigh the reasoning instead of the tone.

A system that can’t explain how it reached a conclusion is asking to be trusted on confidence alone, which is the opposite of honest.

Honesty also means resisting the product incentives that push the other way. A hedge reads as weakness in a demo; “I’m not sure” feels like a worse experience than a confident guess, right up until the guess is wrong and costs someone.

Honest design closes the gap between how sure a system sounds and how sure it actually is — and usually it’s the move that tests slightly worse and serves the user better.

Designing for AI

• Cut the persona. Copy that implies the system thinks or believes overstates what it can actually do — capability inflated and dressed up as friendliness.
• Build an abstain path. A system that never says “I don’t know” is guessing some of the time and hiding it — and the guess that sounds certain is the one that does the damage.
• Flag uncertainty inline. A blanket “AI can make mistakes” footer protects the company, not the user; uncertainty only helps where it sits next to the claim it qualifies.

Good Design Is Long-Lasting

Rams designed against fashion so his products wouldn’t look dated a few years on. Designers under the opposite pressure: models, providers, and best practices turn over every few months, and anything wired tightly to this quarter’s capability ages in a hurry.

Long-lasting design separates the durable part from the volatile part — and the volatile part is whatever model you’re calling under the hood.

A workflow tuned to one model’s quirks — its prompt format, its strengths, the way it phrases things — can break when the provider ships an update you didn’t ask for, which seems to happen every other week. That is the AI equivalent of designing for fashion.

The durable layer is almost always the human one. The user’s goal, their data, and their trust outlast any model generation, and a design anchored there can absorb the turnover underneath.

Build the experience so the engine is replaceable, and the churn becomes maintenance instead of a redesign.

Guardrails are part of that durable layer, and they age badly when they’re scoped only to today’s fast fashion. Build them so they can tighten without a rebuild: a kill switch you can actually reach, audit trails you already keep, human checkpoints you can move as the stakes change. Guardrails designed for this quarter’s model are usually the first thing in the system to expire.

Designing for AI

• Abstract the model. Bind your product to one model and you inherit its every change — and the model is the single part of the stack most certain to be replaced.
• Version the prompts. The model can shift under you without warning, and with no way to notice, your users find the regression before you do.
• Layer the guardrails. Models keep gaining capability and the rules keep tightening, so guardrails scoped to today’s system are the first thing to expire — they have to be able to move on their own.

Good Design Is Thorough Down To The Last Detail

Nothing should be arbitrary, Rams insisted; the care in the details is how you show respect for the person using the thing. AI makes that harder precisely because it’s probabilistic. The demo path looks clean while the edges quietly fall apart — the odd input, the wrong format, the small slice of answers that are confidently, the hallucinations, completely off.

It also means the quality of the answer is now a design concern, not someone else’s. In a traditional product, design stops at the interface and the content behind it belongs to another team.

In AI, the answer is the product — its accuracy, its completeness, its tone — so caring about whether the output is actually good is part of the design work, not a handoff after it.

A polished frame around a mediocre answer is still a mediocre product. Probability changes what “finished” even means. With a deterministic interface, if the button works, it works. With a model, the same input can succeed and then fail, and the failures cluster in the inputs you didn’t think to test.

The polish you see in a demo says almost nothing about the experience at the tenth percentile. The tail is the product.

This is unglamorous work, and it’s where respect for the user actually shows up. Designing the empty result, the malformed output, the confidently-wrong answer, the path back from each — none of it demos well, all of it decides whether people trust the thing.

Rams’ insistence that nothing be arbitrary lands hardest exactly where the model is least predictable.

Designing for AI

• Keep an eval set. The happy path will always look fine, so the only honest measure of quality is what the system does on its worst inputs, release after release. Generate them synthetically so at least you have a starting point.
• Design the failures. With a probabilistic system the failure paths aren’t rare edge cases, they’re a routine part of the experience — and that’s where trust is decided.
• Set the quality bar. A good average hides the tenth-percentile experience, and it’s the tail, not the mean, that decides whether people keep trusting the product.

Good Design Is Environmentally Friendly

Back in the 1970s Rams argued that thoughtless consumption was over and design should conserve resources. The principle has teeth again. Running these models has a real and growing footprint: the IEA projects data-centre electricity use will roughly double by 2030, to around 945 terawatt-hours — close to 3% of world demand — growing about four times faster than electricity use overall.

Designers sit closer to this than it seems.

Defaulting every interaction to the largest available model, re-running a generation because it’s easier than caching, calling a model on every keystroke — these are design decisions, and at scale they’re energy decisions too.

Right now the AI Assistants work like flying a 747 down to the corner store for milk. That’s not responsible.

Remember that the cheapest model that does the job well is often the responsible choice and the faster one at the same time. We should design that into the system.

This isn’t a call to sacrifice the experience for a marginal saving; it’s a call to stop treating compute as free. The same restraint that makes a product feel considered — using the minimum that solves the problem — happens to shrink the footprint.

Rams’ point from the 1970s holds: thoughtless consumption is a design failure, not just an external cost.

Designing for AI

• Route by difficulty. Most requests don’t need the largest model, and reaching for it by default spends energy and money on capability the task never uses. Suggest routing by default.
• Cache and reuse. Re-running the same inference because it’s easier treats compute as free, which it is neither for the user’s wait nor for the grid behind it.
• Batch the calls. Firing the model on every keystroke is developer convenience paid for in cost and latency the user never asked to carry.

Good Design Is As Little Design As Possible

“Less, but better.”

Rams’ last and most-quoted principle is the hardest one to honor when capability is this cheap.

AI invites more of everything — more features, more autonomy, more surface to maintain — because adding is easy and taking away takes judgment. The discipline is finding the smallest move that actually solves the problem.

Subtraction is harder than addition, and AI tilts the field toward addition. Products accrete AI surfaces — a generate button here, an autosuggest there — each defensible on its own, collectively a mess no one chose on purpose.

“Less, but better” advocates for restraint, and it’s where the thread running through the other principles surfaces as the spine of a well crafted message.

The restraint behind being innovative (use the new part to remove, not pile on), unobtrusive (stay quiet until wanted), and environmentally friendly (use the smallest model that works) was never separate disciplines — it was this one all along.

Sometimes that’s a single well-placed model call. Sometimes the best version of an AI feature is a sharper default and no model at all. The hardest design decision in this era is often what to leave out.

Designing for AI

• Design the non-AI version first. If a plain default or rule already does the job, the model is just complexity and cost with nothing to justify it.
• Consolidate the entry points. Scattered AI surfaces each look defensible on their own, but together they become a product no one chose to design.
• Keep the user in command. Background autonomy is useful right up to the moment it commits to something the user would have stopped — checkpoints are what keep quiet help from becoming quiet risk.

Conclusion

Read together, the ten share a common theme, and it isn’t aesthetic. It’s restraint — the willingness to subordinate what you can build to what people actually need. That’s the part this moment makes hardest.

When capability is cheap and shipping is fast, every incentive is more: more features, more autonomy, more confident output. Rams’ list is the counterweight — questions to hold up against that pressure before it it’s part of the roadmap.

Steve Jobs, whose company absorbed Rams most directly, put it more bluntly.

Say no. A lot.

He’d just come back to a flailing Apple, was about to cut its sprawling product line down to a handful, and told a room of developers that “innovation is saying no to 1,000 things.”

Restraint wasn’t a limit on the work. It was the work.

The AI era hands you a thousand new things you could say yes to — which is the whole reason the discipline of no matters now.

It’s also worth remembering where these came from — A designer who keeps no screens in his house, who judges the work by whether it serves a human need and not by what the technology can suddenly pull off, who probably doesn’t use AI. At all.

That distance is the thing most AI work is missing.

Dieter Rams avoids computers. His ten rules still fit designing for AI. was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

AI often makes me feel like I am having one big, long crisis. Like I’m in limbo or purgatory, or something. Doom and gloom, and WTF are we going to do? But then there’s the other side. When I see things people are creating, or I create in my own work (like when an agent pulls some magic out of what feels like nowhere), that makes me feel unbelievably excited.

But no matter what side of that yo-yo I’m on, there are some things that I will always believe.

Firstly, (and this isn’t a new or groundbreaking idea) AI companies, and one in particular, are spinning a tale of productivity because it’s all we’ve got left. We have to find something to eke the last remaining gasp of growth out of capitalism. Like licking the spoon or scraping the bottom of your yoghurt pot, AI companies use that productivity narrative so they can grow and so that we all don’t have to face the fact that maybe we’re reaching a stalemate. This isn’t like other times when there’s been a new innovation or invention. This is different because, for the first time, everyone’s put all their big bets on AI. There’s not one of the big companies that has even a handful of bets elsewhere. So for them it’s AI or bust. Which means they need you to believe that too.

There are obviously so many issues with that. But for me, as a Strategist, the one I worry a lot about is this: These same companies were the ones that had us all trade desire for data, for years. Silicon Valley and the surface areas they created through new forms of content and communication pushed us to focus on conversion more than creativity and handed over the keys to dashboards, click-throughs, and page rankings. And because it happened slowly, and because they owned the platforms, none of us really paused to think. We traded the messy art of culture and imagination for the clean maths of optimisation. And, we ignored what Binet and Field warned us about decades ago. That sacrificing long-term brand building for immediate activation is a trap. They must be holding their heads in their hands right now.

Like many of my ex-Nike colleagues, I know firsthand what a total surrender to “data” can do to a brand. I worked at Nike during one of the most turbulent few years of the company. I was lucky enough to have the incredible job of Director, Key Cities. I worked with incredible teams and two very inspiring (female too!) VPs across London and Paris. They had almost 40 years of Nike experience between them. I sat with them in boardrooms as the team argued that it didn’t matter what brand tracking or sales data said; young people, particularly young girls, were turning their backs on Nike. “We don’t see that in the numbers; it’s all anecdotal”, they’d say. How’s all that data doing for you now?

I look at how everyone is using AI today, and I’m convinced that current usage is not the answer to any of this. But that isn’t the technology’s fault. AI is doing exactly what it was built to do. The danger is that because it seems to know everything, we’ve started outsourcing our actual jobs to it.

Our value, especially as strategists, lies in critical thinking and pulling on our training. Because, yes, AI knows a lot. But because of how AI is trained, AI doesn’t know the real world. And it doesn’t know humans as well as you think. And the real world, and the humans who inhabit it, is where our work matters the most.

Recently, I was out for dinner on Exmouth Market. Sitting in the window of a restaurant, I watched as a woman outside smoked 4 cigarettes in a row. And then when I looked around, I noticed that literally everyone was smoking. I felt like I’d time-warped back to the 1990s. Later that evening, I asked Claude what was going on. He told me I was hallucinating. Kind of rich coming from him tbh but whatever. I then mentioned it to a friend who has two teenage children. They told me what they’d recently said about vaping vs. smoking. Vapes smelled like watermelon, looked like candy, were all over TikTok, and you could get away with smoking them in the classroom if you were quick. Vapes were essentially for babies and chickens. Cigarettes, on the other hand, were the undisputed proof of solid gold cool. You had to be a grown-up. You had to be sneaky. Obviously, that PoV isn’t that new, but since the smoking ban and brilliant public campaigns, surely we’d move past it? I took these real-world insights back to Claude. You’re wrong, he told me. The data says otherwise. And there I was, back in that Nike boardroom again.

I think that creatives are building some amazing things with AI. Their workflows become tools, and their creativity explodes. But Strategists? All of this shows we need to think a bit more carefully about AI and about the parts of our world we give over to the machine.

I’ve been working with AI to varying degrees for almost five years now, and the last 18 months have felt like a total warp-speed accelerator. I’m not looking at this from the sidelines; I’m building with it every day.

Right now, the headlines are littered with corporate disillusionment. Every day there is a new company realising they aren’t seeing the massive productivity gains that the AI giants tout. But in my own day-to-day, I am seeing those gains. Not in the cost-cutting way the optimisation opportunists market, but as a genuine unlock for deeper thinking. And I have seen genuine efficiencies, too. I now do more of what I love, and a lot less of what I don’t. And find answers quicker than ever before.

All this makes me realise that a few things can be true at the same time: AI companies have to aggressively hype generic productivity to scale their capital; the way we discuss productivity today (job loss, marginal gains) is inherently wonky; AI in its current enterprise use is being implemented in a profoundly unproductive way; and yet, the tech absolutely can and will give you immense leverage. But only if you know how to architect it properly. And you will only learn that if you experiment with it every day.

It is through all that experimentation and recent tool building that I’ve realised, if we want to protect the integrity of our own thinking, we have to establish a set of personal rules to follow, and our own personal workflows. It’s through your own work and experimentation that you’ll find yours. But for now, if you’d like some, here are mine.

1. Create like a child, edit like a scientist (thanks, Tyler, the Creator)

AI shouldn’t replace the messy, imaginative, and unconstrained spirit of raw creative strategy. The strategy process needs space for childlike imagination, deep curiosity, and expansive exploration. The machine’s role is to step in afterward to help structure, validate, pressure-test, and refine that original spark.

2. Give AI the desk, keep the street for yourself

The machine is trapped in the screen; you are not. Dump all that soul-crushing admin. The time-consuming competitor audits, the category tracking, the baseline data synthesis, and the pattern recognition. Everything that is computational and transactional can be given to the machine. But retain the world. Strategy doesn’t happen behind a monitor, and it certainly doesn’t happen in a dashboard. Deeper strategy is fueled by human-to-human observation — talking to Uber drivers, visiting galleries, interviewing real people, and capturing the physical, sensory, and messy lived experiences that don’t fit into clean data sets.

3. Inject your personal creative code

If you prompt an agent with the average, it will return the average. When working with AI, and especially when building agents, you must hardcode a foundational set of values, instructions, and constraints that define its creative code. Every agent in your ecosystem must call upon this shared document outlining your specific tastes, your creative beliefs, what you like and dislike, and exactly what “good” looks like to you. Not to the average. Think of it as codifying a cultural standard so the machine knows exactly what level of strategic rigour you expect it to return.

4. Steal from the engineers

The future of strategy isn’t found in looking at what other strategists are doing; it’s found by observing the tools, workflows, and frameworks being built by the engineering pioneers and research scientists at the frontier of AI. And translating these into creative fuel. Look outside your discipline to find the infrastructure for your own thinking. Borrow their automated research architectures, such as using automated engines built for code validation, and re-engineer them to continuously run hypothesis tests, loops, and refinements against strategic creative outputs. GitHub might feel weird at first, but I promise you it’s the best place to start.

5. Build a compounding ecosystem

Stop treating models like search with a personality. AI is at its best when used as a system architecture, not a conversational companion. Command a multi-agent team, build your own tools on demand and have them work together — like alternative narrative agents collaborating with Figma deck builders, or research agents working with interview creator tools. Then, ensure you co-evolve with it. Every real-world observation or unique human tension you bring back from the field must be fed back into your workspace like a knowledge graph. This transforms a generic model into a living ecosystem, compounding institutional memory of exactly how you uniquely see the world and the decisions you make.

6. Human judgement is the final line of accountability

AI does not give you answers; it gives you outputs for your consideration, interrogation, and judgment. You are the original thinker; the machine is not. As the reality of the work dictates: Claude is brilliant, but he can’t get fired (yet). We can. The human strategist always owns the final accountability for the inputs and the outputs.

So, where does this leave us?

It’s easy to get trapped in the anti-AI paralysis. Don’t. But do not surrender to its generic usage either. If you use these tools simply to write your decks faster, or to find you facts and futures, you are actively participating in the machine-replacing-humans narrative. Your job was never to sit behind a desk finding the average; your job was always to get out into the real world, find the friction, and code your own taste into the system.

And now, your job also includes daily experimentation. Build your own tools. Design your own systems. Work at it every single day until you are commanding the ecosystem rather than just chatting with a prompt box. And most importantly, don’t forget to have fun. It’s easy for this to all feel heavy. And get caught up in the corporate hype machine. But if you play and use the curiosity you have innately in you, who knows what you will create.

Originally published at https://msbelt.substack.com.

Strategy in the age of the machine was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

I was working with a client I’ve known for over fifteen years when they casually mentioned that they had asked ChatGPT to review the website design I’d recently delivered. To be clear, we have a strong relationship, and I don’t think there was anything malicious behind it. If anything, the exercise seemed more exploratory than evaluative. Still, for a brief moment, I felt oddly insulted. After all, I’ve spent decades developing expertise in design, and now I found myself discussing my work alongside the opinions of a machine that learned many of its design principles by consuming the very internet that web creators and designers like me helped build.

The feedback itself wasn’t terrible. Some of the points were reasonable. Others, in my opinion, would have created new problems while solving old ones. During our discussion, I acknowledged the useful observations, explained where I disagreed, and offered alternative solutions. Ultimately, the client went with my recommendations.

But this was not the first time it happened.

In fact, several other clients have used AI in some capacity to either evaluate work I had produced or generate suggestions. The first few times it caught me off guard. Not because I was surprised it happened, but because I was surprised I hadn’t anticipated it sooner. The moment you stop and think about it, the behavior is almost inevitable. If clients now have access to a tool that can instantly produce a second opinion, why wouldn’t they use it?

For designers, this creates a strange new reality. It is already difficult enough trying to determine how AI fits into our own workflows. We are still figuring out when to trust it, when to ignore it, and where it genuinely adds value. Now we must also navigate clients and stakeholders using those same tools, often with significantly different levels of understanding about how the systems actually work. In many cases, the conversation is no longer between a designer and a client. It is between a designer, a client, and a machine sitting invisibly in the corner offering unsolicited commentary.

Oddly enough, AI-generated design feedback itself is not particularly controversial. Many designers already integrate AI into their workflows to critique layouts, evaluate accessibility, identify usability concerns, or brainstorm alternatives. I even use it in the classroom. Some of my assignments have my students use guided prompting exercises to receive feedback on certain projects. When the prompts are carefully structured and the objectives are clear, AI can be remarkably useful at surfacing issues that a designer might overlook.

The difference is that those situations are controlled. The designer understands the limitations of the tool and the prompts are intentional. The feedback is filtered through experience and context.

Client use of AI is a different animal entirely. Anyone who has worked in design knows that managing the client relationship is often as challenging as the design work itself. Good designers spend an enormous amount of time translating expertise into language that non-designers can understand. We explain tradeoffs, justify decisions, balance business goals, user needs, technical constraints, and aesthetics. AI now inserts itself directly into that process, often presenting its recommendations with the same confidence regardless of whether they are insightful, superficial, or completely wrong.

The real challenge is not that clients will receive bad feedback. The challenge is that they may receive plausible feedback. Bad advice is easy to dismiss. Plausible advice is far more dangerous because it requires analysis. Every suggestion becomes another conversation and every recommendation becomes another decision to unpack. The designer is no longer just defending design choices — they are increasingly being asked to defend them against an endlessly available synthetic consultant that works for free and never sleeps.

The irony is that this development may ultimately make expertise more important rather than less. When everyone has access to infinite suggestions, we find ourselves living out a modern spin on the attention economy — where data is infinite, but the scarce resource becomes human judgment. AI can generate observations, identify patterns, and even produce surprisingly competent critiques. What it cannot do particularly well is determine which observations matter most within a specific business, audience, budget, timeline, or organizational context. And at the end of the day, someone still has to decide.

So how should designers respond?

• First, don’t take it personally. Clients are not necessarily questioning your expertise — most are simply using the tools available to them. Seeking a second opinion from AI is rapidly becoming as normal as searching Google or reading online reviews.
• Second, ask to see the prompts and feedback. Understanding what the AI was told often reveals why it produced certain recommendations. The quality of the output is usually a reflection of the quality of the input.
• Third, treat AI feedback the same way you would treat feedback from any stakeholder. Evaluate the idea itself rather than its source. Some suggestions will be useful. Others will be irrelevant. Most will fall somewhere in between.
• Finally, recognize that part of the designer’s role is evolving. Increasingly, we are not just creating solutions. As the industry shifts, professional creative roles are moving away from pure production and toward high-level curation. We are interpreting, contextualizing, and filtering an overwhelming amount of information generated by both humans and machines. The designer is becoming less of a producer and more of a translator between competing perspectives.

Perhaps this is simply the next stage of the profession. Many designers are worried that AI would replace them altogether. Instead, we may discover that our new challenge is something far stranger — managing clients and stakeholders who have acquired a tireless digital design intern with unlimited confidence and no accountability. The machine now has a permanent seat at the table. The question is not whether it belongs there. The question is whether we can learn to conduct the meeting while it keeps whispering in everyone’s ear.

Don’t miss out! Join my email list and receive the latest content.

AI has become the third wheel was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

The head of Claude Code at Anthropic, Boris Cherny, recently made a statement that deserves to be read slowly. “I don’t prompt Claude anymore,” he said. “I have loops running that prompt Claude and figuring out what to do. My job is to write loops.” The person building Anthropic’s flagship coding agent does not talk to AI. He designs systems in which AI talks to itself , or ‘self-referential control loops’ that govern what Claude does and how it corrects itself without a human in the cycle. This is the transition, he says, that will define the rest of the year.

“I don’t prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops.” —Boris Cherny (Head of Claude Code, Anthropic).

And we can see this pattern everywhere. Something significant is happening across the AI engineering community, where teams and practitioners are building real systems… I mean the people putting actual code in production. What I’ve observed over recent weeks is that all of them are arriving at the same structural insight through completely independent paths (an example of equifinality). Indeed, they are using different but similar vocabulary, working at different scales, and serving different audiences, but they are describing the same conceptual architecture.

That architecture has a name, and is far from new. It’s not a framework, a product, or a methodology. It’s actually a set of scientific principles, a body of knowledge formalised in the late 1940s by Norbert Wiener and extended through the 1950s and 1970s by W. Ross Ashby, Gordon Pask, and William T. Powers.

My argument here is that modern practitioners should be leveraging the original and stated concepts, rather than using heuristics to rediscover them. As we shall see, it was intentional that the connection was forgotten, but I believe it’s now time to formally rediscover them and apply them.

Those original concepts are from the field of cybernetic, and agentic systems are already rewarding those who apply it in practice.

A short primer on cybernetics

Cyber-what? Indeed, I hear you, it’s been a niche topic over the last few decades, so it demands a proper primer. Cybernetics is the science of goal-directed systems — of how any system, biological, mechanical, social, or designed, pursues a purpose and corrects its own behaviour when it drifts from it. The word comes from the Greek kubernetes (“the steersman”) meaning, not the one who rows but who holds the course. Norbert Wiener formalised the discipline in 1948, when observing that the same principles govern anti-aircraft predictors, nervous systems, and markets. It is not a technology. It is a vocabulary of principles that operate above any specific domain.

The foundational mechanism is the control loop. Every purposive system holds a reference state, or a specification of what it is trying to maintain or reach. It continuously compares that reference against its perception of the current state of the world. When a discrepancy is detected, it generates corrective behaviour to close the gap. William T. Powers (1973) showed this is the universal mechanism underlying all purposive human behaviour, not just in machines, but in every goal-directed act a person performs. You use it all the time to regulate your temperature, your appetite, or keep your car on the road. And one more thing is true, that without a reference state, there is no comparator. Without a comparator, there is no control. There is only execution.

Feedback is the signal that makes correction possible. Negative feedback reduces discrepancy, for example, in the thermostat detecting that the room is too cold and turning the heating on. Positive feedback amplifies a signal, which is the flywheel that accelerates growth. Both have their place, but a system running positive feedback without a balancing negative loop will eventually collapse.

Variety, as defined by W. Ross Ashby (1956), is the number of distinct states a system can occupy. Ashby’s Law of Requisite Variety states that a regulator must possess at least as much variety as the system it is regulating. A system facing more complexity than its response repertoire can match will fail, but not because it lacks intelligence, but because it lacks variety.

Homeostasis is the regulated maintenance of a stable internal state against external disturbance. The goal is not a fixed destination but a dynamic equilibrium. Stafford Beer (1979) extended this to organisations, arguing that a viable system is one that can maintain its own viability through internal regulatory mechanisms. His POSIWID principle (“the purpose of a system is what it does”) cuts through stated intentions and asks what the system demonstrably produces. If the observed behaviour diverges from the stated purpose, the system is regulating toward the wrong objective.

Finally, Gordon Pask (1968) introduced the Phase Space: the full trajectory of states a system moves through over time, not just its current configuration. A system without memory of its trajectory cannot use that history to constrain its next state. A system with it compounds experience into capability.

These six basic concepts (reference state, feedback, variety, homeostasis, POSIWID, and Phase Space (along with many others)) are the ideas through which the following should be interpreted.

Tom Blomfield at Y Combinator: the organisation as viable system

The most operationally direct statement of the convergence comes from Tom Blomfield’s YC batch talk, published 19th May 2026, in which the General Partner and Monzo co-founder argues that most founders are building AI wrong — adding it on top of existing hierarchical structures rather than reconceiving the company itself as a set of recursive, self-improving loops. Blomfield’s core example is a YC internal query agent that partners could use to ask questions about founder meetings and history; useful but unremarkable until a monitoring agent was placed on top of it, watching every query, tracking successes and failures, and — when queries failed — diagnosing why overnight, writing the fix, opening a pull request, having a separate review agent check it, and deploying it before the next morning, so that the same query that failed the night before succeeded without any human intervention. He describes this as his “holy shit moment” — the recognition that the system had regulated itself back to a functional state without being told to.

In cybernetic terms, what Blomfield witnessed was Stafford Beer’s Viable System Model in spontaneous operation: a monitoring function detecting discrepancy, generating corrective behaviour, and closing the loop on a cycle that previously required a human coordinator to complete.

His framing of traditional companies as “Roman legions” (information flowing upward through hierarchies, commands flowing downward through management chains) is a precise description of a variety-attenuation architecture designed for a specific communication constraint; his argument that AI dissolves this structure is the argument that Ashby’s Law now operates at organisational scale. When an AI system can match the variety of operational disturbances that middle management previously absorbed, the human attenuation layer is no longer necessary for regulation.

His closing observation that “software is ephemeral, context is valuable” names the same distinction that Murag (see below) makes through memory architecture: what persists and compounds in value is not the implementation but the accumulated organisational knowledge that defines what the system should do, the reference state against which all subsequent regulatory behaviour is calibrated.

Mahesh Murag at Anthropic: memory as Phase Space, Dreaming as second-order homeostasis

The most architecturally complete statement of what the convergence requires at the infrastructure level comes from Mahesh Murag’s session at Code with Claude San Francisco, published 8th May 2026, in which the Anthropic product manager who built the Model Context Protocol argues that memory is the next primitive — the missing building block that turns agents from single-session executors into systems that accumulate capability over time.

His argument is a direct statement of Pask’s Phase Space principle: a system without memory begins each session in the same state regardless of prior trajectory and cannot be state-determined, because history is not accessible to it; with memory, the current state encodes the trajectory, and what the agent has learned compounds rather than evaporates.

The design decision that most clearly reflects cybernetic thinking is not that memory exists but how it is modelled: rather than imposing fixed schemas, Anthropic lets Claude manage memory as a plain-text file system the agent organises for itself, applying Ashby’s Law at the memory-architecture level — a fixed-schema store whose structural variety is less than the variety of the agent’s operational experience will systematically fail to encode what matters.

The feature Murag presents as most novel is Dreaming: a background consolidation process that runs outside the agent’s normal work path, reads recent sessions alongside existing memory, removes duplicates and stale information, and surfaces patterns no single session had enough perspective to detect — a second-order homeostatic mechanism, analogous to biological memory consolidation during sleep, that maintains the quality of the regulatory architecture itself rather than the quality of any individual task output, and whose costs are paid once while benefits compound across every subsequent session.

Most significantly, Murag identifies shared memory across agent swarms as the prospect he finds most compelling. Imagine if hundreds of agents in parallel contributing trajectory observations to a collective state representation, making cross-agent patterns visible that no individual agent could detect. And from the cybernetic perspective, this is arguably the mechanism by which Wiener’s (1948) made his original insight about shared error signals increasing collective regulatory capacity. This ‘new’ discovery by Murag shows how Wiener’s model extends, 78 years later, to fleets of AI agents building a shared understanding of their environment over time.

Daisy Hollman: “The secret isn’t a better model — it’s tighter feedback loops”

At Anthropic’s Code with Claude conference in London on 22nd May 2026, Daisy Hollman (a Member of Technical Staff at Anthropic) gave a workshop titled “Beyond the Basics with Claude Code.” One of the central claims, delivered at 14:30, is a precise statement of a cybernetic principle to have emerged from the practitioner community this year (so far).

The secret to doing great work with Claude Code is not a better model. It is tighter feedback loops. — Daisy Hollman (Member of Technical Staff, Anthropic).

This is Wiener’s foundational argument from 1948, restated for a 2026 engineering audience without the Wiener. Wiener’s original observation was that the performance of a purposive system — any system trying to achieve a goal — is determined by the quality of the feedback mechanism that corrects its behaviour, not by the raw power of its components. An anti-aircraft predictor that cannot receive information about where its shells are landing will not improve, regardless of how sophisticated its ballistic calculations are. A Claude Code agent that cannot receive structured information about the quality of its outputs will not improve, regardless of model capability.

Hollman makes the same point through a different route: context windows have not grown in over a year, and she does not expect that to change. While raw capability has scaled, the constraint is informational — specifically, the selection of what goes into the fixed box. She describes hooks as “red squigglies for agents” — small corrections injected at the moment of a mistake rather than caught later in review. This is the principle of early-cycle feedback: detect discrepancy at the earliest possible point in the loop rather than post-hoc. Powers’ Perceptual Control Theory (1973) is built entirely on this principle. The value of a feedback signal is inversely proportional to the delay between error and correction.

Her framing of memory as a “context engineering primitive” (30:00) is the second cybernetic insight in the talk. Context is not a container — it is the agent’s perception of the current state. What goes into the context defines what the agent can detect, and what it can detect defines what discrepancies it can act on. This is not a memory management problem. It is a perception architecture problem. Powers’ hierarchy is explicit: a control system can only regulate what it can perceive. Hollman has arrived at the same conclusion from the engineering side.

Nick Saraev: the DOE framework as control loop architecture

Nick Saraev, who runs a Claude-focused channel and Skool community, published a course on advanced Claude Code workflows in which he separates the agentic workflow into three components: Directives (i.e. SOPs , or the standing operating procedures that define expected behaviour), Orchestration (the AI “brain” that directs activity), and Execution (deterministic Python scripts that carry out specific tasks). He frames this as a way to stop AI from hallucinating and start making it reliable for business.

The DOE framework is a control loop in practical form. The Directives are the reference state: a specification of how the system should behave. The Orchestration layer is the comparator and action generator: it evaluates current behaviour against the Directives and selects the next action. The Execution layer is the effector: the mechanism that produces observable changes in the world. The separation is structural rather than stylistic because a system in which the agent that specifies the reference state also generates the action, and evaluates the output has collapsed the comparator into the actor. Saraev has independently discovered that this produces hallucination and unreliability, and that the remedy is separation of function.

The self-referential loop he describes is the application of this structure recursively: the Orchestration layer continuously compares the state of execution against the Directives and generates corrective behaviour until the condition is satisfied. This is exactly what Anthropic’s /goal command implements at the platform level — with the added cybernetic refinement that the evaluator is a separate model from the one doing the work. The comparator must not be the same system as the actor.

Luke Alvoeiro at Factory: serial execution, validators, and adversarial verification

Luke Alvoeiro’s talk at Code with Claude San Francisco on 6th May 2026 (“The Multi-Agent Architecture That Actually Ships”) makes the most structurally complete argument of any practitioner presentation in the current period. His taxonomy of five frontier multi-agent strategies and the three-role production system he describes (orchestrator, workers, validators) map precisely onto a distributed control loop.

The orchestrator holds the goal and decomposes it into subtasks. The workers generate outputs. The validators are comparators: their function is to detect discrepancy between the workers’ outputs and the expected standard, using what Alvoeiro calls “validation contracts” and “adversarial verification.” The adversarial element is significant — the validator is explicitly designed to find discrepancy, not to confirm adequacy. This is the design of a sensitive comparator. In Powers’ model, a comparator calibrated to find discrepancy rather than confirm completion produces tighter loops and faster correction.

Alvoeiro’s argument for serial over parallel execution is also cybernetically precise. Parallel execution generates multiple outputs simultaneously without inter-agent feedback during generation. Serial execution allows each step’s output to inform the next — feedback can propagate through the chain in real time. The case for serial is not about latency. It is about feedback architecture. Tighter loops require sequential dependency.

His third argument — that model selection per role is a “compounding advantage” — is an application of Ashby’s Law of Requisite Variety (1956) at the component level. A validator requires different capabilities than a generator. Assigning the wrong model to a role reduces that component’s regulatory capacity below the variety of the inputs it faces. Over time, this compounds: a validator with insufficient variety will systematically fail to detect certain classes of discrepancy, and those failures will accumulate invisibly in production outputs.

His stated goal of designing systems “that get better with each model generation instead of being made obsolete by them” is the most strategically sophisticated idea in the talk. It is a principle of regulatory architecture over implementation architecture: design the control loop structure so that model improvements increase the system’s regulatory capacity without requiring structural redesign. The value of the system is in the feedback structure, not in the specific components instantiated within it.

The Ralph Loop and recursive self-improvement

The practitioner community has also independently discovered what the Alibaba Cloud engineering blog calls the “Ralph Loop” — a self-referential iterative loop that allows an agent to continuously see its own previous outputs through external state (code, test results, commit records) and iterate until a condition is satisfied. The core mechanism is that the agent’s output becomes part of its next input via the file system and version history. Not simply an “output as input”, but rather it is feedback through an external state representation, which is precisely the mechanism that allows a control system to maintain a model of the world rather than simply responding to immediate stimuli.

Anthropic’s Managed Agents features (announced at Code with Claude 2026) formalise this. Effectively, outcomes let you define success criteria so agents can iterate and improve over time; Dreaming allows Claude to recall previous sessions and build on past work. These are examples of homeostatic mechanisms operating at different time scales. Outcomes is a per-task comparator, and Dreaming is a cross-session memory update, or the accumulation of experience across the through-states of the agent’s operational life-span.

The MindStudio analysis of compounding knowledge loops describes the same structure. By pairing session lifecycle hooks with an automatically updated knowledge base, you can create a Claude agent that genuinely gets smarter over time . This is where each session leaves the agent better equipped for the next, and therefore this is homeostasis at the organisational level. The agent’s reference state is improving across sessions, and the feedback architecture ensures that improvements are retained and compounded rather than lost at session end.

The self-evolving COO

The concept of a self-evolving organisational agent — the “self-evolving COO” — extends the same architecture to the level of business operations. ServiceNow’s Project Arc, announced at Knowledge 2026, is described as “a long-running, self-evolving autonomous desktop agent… that connects natively to enterprise systems.” Unlike standalone AI agents, Project Arc connects natively to enterprise workflow context and governance from ServiceNow AI Control Tower.

The research taxonomy of self-evolving agents is more precise about the architecture. Emergentmind’s synthesis of self-evolving agent papers identifies the key axes: what to evolve (model parameters, prompts, memory, tools, workflow graphs, or agent roles); when to evolve (intra-task via test-time reflection, or inter-task via evolutionary search across episodes); and how to validate evolution (dual audits, ablation studies).

This is the design of a second-order homeostatic system: a control loop that maintains not just a performance state but the capability to maintain that state as conditions change. Stafford Beer described this as the viable system requirement: a system that can adapt its own regulatory architecture in response to environmental change is qualitatively more robust than a system that can only regulate within a fixed architecture.

The great remembering

The convergence documented in this article is not simply independent rediscovery. I argue, that it is actually the return of a suppressed origin. To understand why, it is necessary to know a piece of history that most practitioners building AI systems today might not have encountered.

In the summer of 1955, John McCarthy (the man who coined the term “Artificial Intelligence”) was planning what would become the Dartmouth Conference of 1956 , the founding event of the field we now call AI. As he prepared the proposal, he faced a naming decision. The natural choice was “cybernetics” because it was already in use, already established, already the vocabulary for precisely the problems he intended to address. As we’d seen earlier, Norbert Wiener had coined the term, defined the field, and written its foundational text. Cybernetics was the right word.

McCarthy chose a different one. His own explanation survives in his Stanford archives: “one of the reasons for inventing the term ‘artificial intelligence’ was to escape association with ‘cybernetics.’ Its concentration on analog feedback seemed misguided, and I wished to avoid having either to accept Norbert Wiener as a guru or having to argue with him.” mexc

The decision was partly intellectual because McCarthy was interested in digital computing rather than analog feedback, but it was also partly personal. Wiener was, by multiple contemporary accounts, a difficult character. He was often possessive of his ideas, a poor listener, prone to dominating conversations in fields not his own. McCarthy didn’t want cybernetics’ most prominent figure to claim ownership of the new conference. So he invented a new name, and a new field, and a new vocabulary that pointed away from the one Wiener had built.

As Punya Mishra observes, this choice had profound and far-reaching consequences. Unlike “cybernetics,” which was relatively neutral, “intelligence” carries significant philosophical and emotional weight. It implies a direct comparison to human cognition and has led to decades of debate about whether machines can truly “think.” The anthropomorphic framing shaped research directions, public expectations, and funding priorities in ways that a more neutral, systems-oriented vocabulary might not have. Every subsequent debate about whether machines are conscious, whether they “hallucinate,” whether they will surpass human intelligence. These are arguably the downstream effects of a naming decision made partly to avoid an argument with a difficult colleague.

The practical consequence for engineering was subtler but equally significant. By departing from cybernetics, the field departed from its vocabulary of feedback, variety, control, homeostasis, and reference states. The concepts did not disappear but were absorbed into control theory, systems engineering, biology, and organisational science, where they continued to develop. But the community building AI systems lost the shared language that connected them to those concepts. Each generation of engineers has had to independently derive, from production failure, the principles that cybernetics had formalised by 1956.

This is what makes the convergence documented in this article more than an interesting coincidence. Daisy Hollman’s “tighter feedback loops,” Luke Alvoeiro’s “validation contracts,” Mahesh Murag’s “Dreaming,” Tom Blomfield’s “recursive self-improving loops,” Boris Cherny’s “my job is to write loops” — none of these practitioners are drawing on Wiener, Ashby, Powers, or Beer. They are arriving at the same principles through the pressure of production systems that fail when those principles are absent. The field is finding its way back to its own forgotten foundations. In this case not through scholarship, but through the irresistible logic of building systems that are required to work.

I believe that McCarthy was right that cybernetics and digital AI were not identical. Who knows what might have happened if he had not drawn a distinction. Might we have had AI earlier, or later, or at all? But one thing is more certain. The principles cyberneticians established such that purposive behaviour requires feedback, that regulators must match the variety of what they regulate, that the purpose of a system is what it does — operate regardless of whether the substrate is analog or digital, biological or computational. They are not properties of a technology. They are properties of any goal-directed system. And systems that ignore them fail, whatever they are called.

The forgotten patterns

So what can we observe in this article? We can see how across a Skool course, multiple Anthropic staff workshops, a Factory engineering talk, multiple platform feature announcements, a research taxonomy, and many more on a week-on-week basis — the same structural pattern recurs:

A system generates output. Something separate from the generator evaluates that output against a reference. If discrepancy is detected, the loop continues with the evaluator’s finding as input to the next cycle. The loop terminates when the reference state is satisfied or a bound is reached.

What makes this convergence particularly striking is that it is happening simultaneously across disciplines, not just across practitioners. On 11th May 2026 — the same month as the talks documented in this article — researchers at Singapore Management University, Hong Kong Polytechnic University, Nanyang Technological University, A*STAR, and Shanghai AI Lab published “The Agent Use of Agent Beings: Agent Cybernetics Is the Missing Science of Foundation Agents” (arXiv:2605.10754, echoing a paper written by Weiner about the Human Use of Human Beings). Their conclusion was identical to the one this article draws from the practitioner evidence: engineering practice has converged on useful primitives assembled by empirical trial and error rather than from first principles, and cybernetics provides the missing theoretical scaffold.

A parallel paper from the control engineering community — Eslami & Yu’s “A Control-Theoretic Foundation for Agentic Systems” (arXiv:2603.10779, 2026) — arrived at the same architectural conclusions from a dynamical systems perspective, developing a five-level hierarchy of agent autonomy grounded in closed-loop control theory. Academic computer science, control engineering, and production practice, working entirely independently, converged on the same point in the same month. We can see that the equifinality is now demonstrable across disciplines.

And as we have seen, this is Powers’ Perceptual Control Theory (1973), Wiener’s negative feedback loop (1948), and Ashby’s regulatory system (1956). It is the universal architecture of purposive behaviour, described formally in the mid-20th century and is what I believe is being independently rediscovered by practitioners in 2026 because it is finally possible to build systems complex enough to require it. (To be fair, society was already a complex enough system to require it, but humanity tends to suffer from the Solomon effect. Now we need to solve it to use AI effectively).

The practical implications of naming the convergence are not academic. Teams that understand the underlying cybernetic principles can:

1. Design the comparator first — before the agent, before the prompt, before the tool selection. The reference state specification is the primary design decision; everything else serves it. This is a forcing mechanism for you to think through what you’re actually trying to change, or what value you’re optimising.
2. Instrument the feedback loop explicitly. Not as a monitoring afterthought, but as the load-bearing structure of the system. Hollman’s 80% prompt cache hit rate target is an operationalisation of this: if you cannot measure cache efficiency, you cannot optimise context selection, and if you cannot optimise context selection, you cannot tighten the feedback loop.
3. Apply Ashby’s Law as a design constraint — before specifying each agent’s role, audit the variety of the problem space it will face. A validator whose response repertoire does not cover the range of failures the workers can produce is not a comparator. It is a checkbox.
4. Distinguish homeostatic mechanisms by time scale — per-task loops (Saraev’s DOE, Alvoeiro’s validators), per-session persistence (hooks, Dreaming), and cross-session evolution (compounding knowledge loops, self-evolving agents) are all negative feedback mechanisms operating at different temporal resolutions. They compose, but they must be designed separately.

And now it appears that the engineering community has arrived, through applied practice, at a conclusion that the cybernetic community reached over 70 years ago: that purposive systems require comparators, and comparators require reference states, and reference states must be specified before anything else is built.

The name for the science that established this is cybernetics. And it might just be the skill the 2026+ job market needs, without even knowing it.

What do you think? Will cybernetics will be the next big in-demand skillset in 2027?

Disclosure: I am an Ex-officio council member of the Cybernetics Society. Massive thank you to my fellow Cyberneticians, and The Cybernetics Society for filling my brain with wisdom for years. My fellow council members Peter Tuddenham (current VP of the Cybernetics Society), and Alan Outten who proof read and provided feedback on this piece.

References

Wiener, N. (1948). Cybernetics: Or Control and Communication in the Animal and the Machine. MIT Press.

Ashby, W. R. (1956). An Introduction to Cybernetics. New York: Wiley.

Powers, W. T. (1973). Behavior: The Control of Perception. Aldine de Gruyter.

Beer, S. (1979). The Heart of the Enterprise. Chichester: Wiley.

Anthropic (2024). Building effective agents. anthropic.com/engineering/building-effective-agents

Hollman, D. (2026). Beyond the basics with Claude Code. Code with Claude London, 19th May 2026. youtube.com/watch?v=tuY2ChJIx48

Alvoeiro, L. (2026). The Multi-Agent Architecture That Actually Ships. Code with Claude San Francisco, 6th May 2026. youtube.com/watch?v=ow1we5PzK-o

Ebert, C. (2026). Notes from Code with Claude 2026. chrisebert.net

Alibaba Cloud (2026). From ReAct to Ralph Loop: a continuous iteration paradigm for AI agents.

MindStudio (2026). Compounding knowledge loop: Claude Code.

Emergentmind (2026). Self-evolving AI agents.

Anthropic Claude Code documentation (2026). /goal command. code.claude.com/docs/en/goal

McCarthy, J. (1955). A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence.

Mishra, P. (2024). Cybernetics or AI? What’s in a name? punyamishra.com

Cherny, B. (2026). Quoted in Guillermo Flor, LinkedIn. June 2026.

Wang, X., Yang, C., Zhao, H., Lin, Z. & Hu, S. (2026). The agent use of agent beings: agent cybernetics is the missing science of foundation agents. arXiv:2605.10754

Davies, D. (2024). The Unaccountability Machine: Why Big Systems Make Terrible Decisions — and How the World Lost Its Mind. Profile Books.

Eslami, A. & Yu, J. (2026). A control-theoretic foundation for agentic systems. arXiv:2603.10779

Koralus, P. (2025). The philosophic turn for AI agents. Mind & Society, 24(2), 563–586.

Blomfield, T. (2026). How to build a self-improving company with AI. YC Root Access, 19th May 2026. youtube.com/watch?v=t-G67yKAHBQ

The forgotten science behind self-improving companies was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

Volunteers in a 1960s psychology study were played a tape of a man auditioning for a quiz team. He was dazzling, reeling off answer after answer, until near the end he clumsily knocked a cup of coffee over himself. Oddly, the people who heard the spill warmed to him more than those who heard a clean run. The blunder, not the brilliance, was what tipped him over into likable. Psychologists call it the pratfall effect, and it is bad news for anyone paid to make things look perfect.

The design industry, after all, is sanding every last coffee-spill out of its work, and doing it with real discipline.

Spend an afternoon scrolling through its recent output. Everything is good. The corners curve to the same friendly radius, the gradients melt without a seam, the copy reads like the copy on every other site. It is all competent and tidy and precisely where it should be. The longer you look, the harder it gets to remember any of it.

This sameness is a survival strategy. Teams are leaner than they were, deadlines arrive sooner, and the same stretch of hours now has to cover work once shared between several people. Under that kind of strain, a visible mistake feels like a luxury few can afford. No designer wants to be the one whose wonky kerning slipped through to the client, so it gets smoothed away fast, and the version that ships is the one least likely to land anyone in trouble. Defensible beats interesting when your job might be next.

And flawless is wonderfully defensible. Nobody ever got fired for a tidy grid. It gives a jittery team something solid to stand behind when the money is short and the feedback is brutal. You can see exactly why a trade under pressure reaches for it. The instinct is sound. The timing is a calamity.

Why flawless stopped being impressive

For most of this trade’s history, polish was proof that someone had bothered. A crisp line meant a steady hand and a great deal of practice. A seamless layout meant somebody had wrestled with it well past midnight. The smoothness was a receipt, and we stopped noticing we were even reading it.

That receipt is now worthless. A generative model will turn out the same line and composition, along with dutifully on-brief text, in the time it takes to type the request, and charge a pittance for the favour. The craft is still there. It has just gone invisible, because the finished thing looks identical whether it cost a fortnight or four seconds.

An irony is folded into that speed. The model can summon that perfection only because it learned from millions of human-made pieces, each one the product of the very effort it now appears to skip. Its flawlessness is borrowed, congealed from human labour that no longer shows up anywhere on the page.

So here is the design industry, sprinting flat out toward the one quality it can no longer be paid for. The faster it runs, the more its work comes to resemble the free, instant, machine-made version it is trying to rise above. Perfection used to mark out a human who cared. Now it is the shortest path to being mistaken for a machine that didn’t.

Which leaves an awkward question hanging. If a perfect finish no longer proves anything, what does? Inconveniently for everyone, the answer is the very thing the field has spent the past few years teaching itself to delete.

The competence catch

That coffee-spill experiment dates to 1966, and its finding was not simply that mistakes are charming. It came with a sharp condition attached.

When the clumsy candidate was someone who had just answered most of the questions correctly, the spill made him more appealing. When the same coffee went flying from a candidate who had fumbled the quiz, it cost him. The blunder only flattered the people who were already good. On everyone else, it just confirmed the doubt.

That condition is the part most people skip, and it is why “embrace imperfection” is such treacherous advice on its own. A rough edge on confident, accomplished work reads as character. The same rough edge on a weak piece reads as exactly what it is. Imperfection is not a substitute for skill. You earn the right to show it.

What we were paying for all along

The pratfall explains why a flaw can flatter. It does not explain why we are so drawn to human creations in the first place. For that, look at what happens when people make something themselves.

Back in 2012, three behavioural researchers had volunteers assemble plain IKEA storage boxes, fold origami and build Lego, then asked what the results were worth. The builders consistently valued their own slightly shonky attempts far above what anyone else would pay, and assumed strangers would see the genius too. The team behind the study named it the IKEA effect and found it held only when the task was finished. Abandon the flat-pack halfway, and the warm glow disappears.

The effort, then, was never incidental to how we valued the object. It was woven into it. A related strand of work makes the point even more plainly. Marketing scholars found in 2015 that simply describing a product as handmade led people to rate it more attractive and pay more for it, a pull they traced to a sense that such items “contain love” from their makers. The same mug is worth more once a person has shaped it.

Notice what these studies are not about. None of them concern how the finished thing looks; they get at what we believe went into it. We were never only buying the line or the layout. We were buying the hours, the judgment, the proof that another person had spent a part of themselves on our behalf. Automation is superb at the line and the layout. It spends none of itself.

Telling people a machine made it

If the human trace is what we are paying for, then being told a machine made something should change its value, even when the object itself is identical. It does, and the size of the effect is hard to ignore.

The clearest evidence sits in a 2023 study from Columbia Business School, spanning six experiments and nearly three thousand people. They showed participants pieces of art, labelling some as human-made and some as AI-made, and often swapped the labels around on identical images so that any difference in response could only come from the tag.

It did a remarkable amount of work. Anything attributed to AI was judged less skilful, less valuable and less deserving of the word “art”, even though more than seven in ten participants admitted they could not tell the human and machine images apart. In one experiment, the AI label alone knocked an estimated 62% off a painting’s value and 77% off how long people guessed it had taken to make. Nothing had changed but the story of who made it.

There is a twist that should interest anyone earning a living alongside a model. Placing a human-made piece directly beside an AI counterpart made it look more creative than the same piece did among its own kind. Contrast with the machine flattered the maker. The researchers suggest artists might do well to invite that comparison rather than dread it.

You can watch this happen to a brand in real time. In late 2024, Coca-Cola remade its much-loved 1995 “Holidays Are Coming” advert largely with generative AI. The result was technically slick yet emotionally hollow, and audiences said so at volume, calling it “soulless” and noting the irony of a “Real Magic” tagline on a film barely any human had touched. Coca-Cola defended the experiment and ran another AI version the next year, to much the same reception. The timing is the point: at the exact moment a campaign is meant to feel warm, the machine-made polish read as the opposite.

The reaction was widely shared. In the same year, a Getty Images survey found that around nine in ten people wanted AI-generated imagery disclosed, and reported thinking less of brands that used algorithms to create images of people or products. “Made by a human” is quietly turning into a selling point. What backs up that label is the evidence a person was here at all: the wobble, the texture, the choice a machine would have averaged into mush.

The mistake that stuck

Sometimes the flaw does more than signal humanity. Every so often it is the better product outright, and design history is full of examples.

In 1968 Spencer Silver, a chemist at 3M, set out to invent a much stronger adhesive. He produced the opposite: a weak, low-tack glue that clung lightly and peeled away without a trace. By the terms of his own brief, it was a dud. He spent years touting it around the company regardless, finding no takers, because nobody could work out what such a feeble glue was for.

The answer arrived in 1974, when his colleague Art Fry grew tired of the paper bookmarks that kept sliding out of his church hymnal. Silver’s useless glue happened to be ideal for a bookmark you could lift and reposition without tearing the page. The Post-it note, now stuck to a few billion surfaces, is a flaw that found its purpose the moment somebody stopped treating it as a defect.

It would be glib to file this under “mistakes are good”, because most lead nowhere, and the rest merely pave the way for what does. Silver’s glue mattered because Fry was paying close attention to spot a use everyone else had missed. That is the harder lesson hiding inside the happy accident. Surprises only pay off in a process with enough slack to notice them, and enough nerve to keep the strange result rather than bin it for missing the spec.

Putting the flaws back in

Slack and nerve are exactly what is in short supply right now. The reflex, when the pressure is on, is to polish everything until it gleams. That made sense when a flawless finish was scarce and costly. The logic barely holds now that it is available to anyone with a subscription and a spare afternoon.

The more useful instinct is the opposite. Leave the fingerprints in. Keep the asymmetry a grid would have straightened, the word slightly too human to be predicted, the texture that survived because someone decided it should. Let the work show that a person was behind it, small, deliberate irregularities and all.

Two warnings keep this honest, both from the research. One concerns competence: a pratfall only rewards what is already good, so this is no licence for sloppiness dressed up as style. The other concerns attention: a happy accident counts for nothing until someone is alert enough to keep it, which takes a process with room to spot a useful slip before it gets tidied away. None of which is an argument against skill. Polish is still hard won, and still worth wanting. What it no longer does is convince anyone to pay a premium.

For years, a flaw was something to fix and forget. It may now be the most valuable thing you can leave in. The feature, all along, was the flaw.

If you enjoyed this, follow me on Medium for more on design, psychology and technology.

References & Credits

Aronson, E., Willerman, B., & Floyd, J. (1966). The effect of a pratfall on increasing interpersonal attractiveness. Psychonomic Science, 4(6), 227–228. https://doi.org/10.3758/BF03342263

Fuchs, C., Schreier, M., & van Osselaer, S. M. J. (2015). The handmade effect: What’s love got to do with it? Journal of Marketing, 79(2), 98–110. https://doi.org/10.1509/jm.14.0018

Getty Images (2024). Building Trust in the Age of AI. https://newsroom.gettyimages.com/en/getty-images/nearly-90-of-consumers-want-transparency-on-ai-images-finds-getty-images-report

Horton, C. B., White, M. W., & Iyengar, S. S. (2023). Bias against AI art can enhance perceptions of human creativity. Scientific Reports, 13, 19001. https://www.nature.com/articles/s41598-023-45202-3

NBC News (2024). Coca-Cola causes controversy with AI-generated ad. https://www.nbcnews.com/tech/innovation/coca-cola-causes-controversy-ai-made-ad-rcna180665

Norton, M. I., Mochon, D., & Ariely, D. (2012). The IKEA effect: When labor leads to love. Journal of Consumer Psychology, 22(3), 453–460. https://doi.org/10.1016/j.jcps.2011.08.002

University of Colorado Boulder (2013). Origins: Post-it Note Adhesive. Coloradan. https://www.colorado.edu/coloradan/2013/12/01/origins-post-it-note-adhesive

The flaw is the feature was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.