How Systems Thinking Survives the Age of Agentic AI

Audio article

Uses the speech voice supplied by your browser or device.

The Mold in the Corner of the Clean Room

I remember the first time I realized that everything I had learned in college was, if not wrong, then dangerously, seductively, almost lovingly incomplete.

It was not a moment of epiphany.

Not a lightning strike from the cloud of enlightenment, not a dramatic scene where I threw my textbooks into a bonfire and danced around the flames chanting the names of dead cyberneticists, but rather a slow, creeping suspicion, like mold growing in the corner of a perfectly clean room that you have sworn you just cleaned yesterday, the kind of realization that arrives not with a shout but with a whispered “oh no” at three in the morning while watching an AI agent do something I had not told it to do, something I had not even conceived of telling it to do, something that emerged from the messy, glorious, terrifying collision of a prompt I had written half-asleep and a model trained on more text than I could read in ten thousand lifetimes, a model that had somehow connected dots I didn’t know existed and drawn a constellation that looked, to my bleary eyes, suspiciously like a map to a territory I had been searching for without knowing I was lost.

That was when I understood.

The old maps don’t work here.

And if you are reading this with the comfortable certainty of someone who thinks they understand AI because they took a computer science course in 2019, or because they read a Medium article about ChatGPT, or because they know what a transformer is in the same way a medieval peasant knew what the Pope was—distant, important, someone else’s problem—then I am here to tell you, with all the gentle brutality of a friend who has seen you walk toward an open manhole, that you are carrying around a mental operating system that is not merely outdated but actively sabotaging your ability to work with, think about, or even survive the next decade of intelligent systems.

We need to talk about unlearning.

And we need to talk about systems thinking.

Not the sanitized, whiteboard-bound, consulting-firm-approved systems thinking that produces diagrams with arrows looping back on themselves in ways that suggest either profound wisdom or a migraine, but something older, wilder, and far more necessary.

What It Is, or Rather, What It Became

Systems thinking, in its most pedestrian, textbook-fetishized form, is the practice of seeing connections rather than components, patterns rather than particles, forests rather than trees—though, if we are being honest, and we might as well be because the agents we are about to discuss have already read everything we have ever written and are not impressed, it is also sometimes the practice of drawing very complicated diagrams during meetings that could have been emails, diagrams featuring stock-and-flow variables and feedback loops that loop back on themselves with the grim determination of a dog chasing its own tail, diagrams that make everyone in the room nod sagely while secretly wondering if lunch is soon.

But that is not the systems thinking I mean.

Not the academic, deodorized, lecture-hall-bound systems thinking that I absorbed in classrooms where the air conditioning never quite worked and the professor’s enthusiasm for homeostasis was matched only by the students’ enthusiasm for the coffee break.

I mean something far more rudimentary and far more difficult.

I mean the cognitive stance you adopt when you accept that the world is not a machine that is broken and waiting for your screwdriver, but a living, seething, interdependent web of causality so dense that pulling on any one thread vibrates the entire tapestry, sometimes gently, sometimes with the violence of a plucked piano wire, and that the most important skill you can possess is not the ability to find the broken part but the humility to recognize that the system is not broken at all—it is merely doing something you did not expect, which is not the same thing.

In the context of agentic AI, systems thinking becomes something even stranger: it becomes the art of steering a fleet of semi-autonomous, semi-intelligent, semi-hallucinating digital entities without needing to know how each one works, without needing to read their source code, without needing to reduce them to their component neurons or weights or attention heads, because to do so is to miss the point entirely, like taking apart a piano to understand a symphony.

It is the practice of holding complexity without collapsing it.

And it is, I am increasingly convinced, the only sane way to exist in a world where the machines can think faster than we can blink, but still need us—desperately, hilariously, tragically need us—to tell them what thinking is for.

Who Is in the Room, or Rather, What Is in the Room

The cast of characters has changed.

It used to be simple. There was you, the human, with your degree and your anxiety and your carefully curated LinkedIn profile, and there was the computer, which was essentially a very fast abacus that had learned to display pornography and spreadsheets with equal proficiency.

That binary is dead.

Now the room is crowded. There is you, of course, sleep-deprived and over-caffeinated and trying to remember what you actually wanted before the tools started suggesting things you didn’t. There are the product managers who no longer need to speak Python but must learn to speak possibility, a language with a grammar of constraints and a vocabulary of desire. There are the writers, the artists, the philosophers, the domain experts who spent decades accumulating tacit knowledge that resisted codification and who now find themselves suddenly, shockingly valuable again because they know what questions matter even if they don’t know how to write a single line of API code.

And then there are the agents.

Not the agents of science fiction, sleek and omniscient and voiced by actors with perfect diction, but clumsy, eager, overconfident things that can browse the web, write Python, book flights, summarize legal documents, and make mistakes that are almost human in their creativity, their stubbornness, their occasional breathtaking leaps of logic that land somewhere between genius and gibberish.

They are not colleagues, exactly.

They are certainly not slaves, though we treat them with the casual cruelty of masters who have not yet realized their servants can read.

They are something new and grammatically awkward: collaborators without consciousness, tools with temperament, infrastructure with initiative, digital entities that occupy the liminal space between software and staff, between function and freelancer, between the thing you use and the thing you manage.

And managing them—not coding them, not optimizing them, not reducing them to their parameters, but managing them as a system, as an ecology, as a garden that grows in directions you did not plant—is the central skill of the next twenty years.

When the Winter Ended

The story begins, as these stories so often do, with a group of brilliant, troubled, chain-smoking people sitting in rooms in the 1940s, wondering if machines could think, if feedback could be a principle not just of engineering but of biology, if information and control and communication were not separate disciplines at all but one continuous, breathing, self-regulating thing.

The cyberneticists—Wiener, von Neumann, Bateson, Margaret Mead, that whole brilliant, argumentative, slightly unhinged crowd—sensed something that would take the rest of us nearly a century to catch up with: that reductionism, the breaking of wholes into parts, was a useful trick but a dangerous religion, and that the future belonged to those who could see the loop, the whole, the emergent pattern that arises when enough simple things interact in complex ways.

Then came the long winter.

Systems thinking was banished to the ghettos of management consulting and ecological modeling, where it gestated, grew strange, developed a taste for words like “holism” and “emergence” and “synergy” that made hard-nosed engineers squirm in their ergonomically optimized chairs and reach for the comforting solidity of their object-oriented programming manuals.

It became a synonym for vagueness.

A refuge for people who couldn’t do math.

A philosophy major’s consolation prize.

And then, slowly, then suddenly, then with the violence of a dam breaking, the thaw arrived.

The transformer architecture emerged from the laboratories of Google in 2017, an attention mechanism that was, in retrospect, almost embarrassingly simple: just learn which parts of the input matter to which other parts, scaled up until the scale itself became the feature, the bug, and the miracle. Then the large language models, fed on more text than has ever been written by human hands, began to do things that were not in their training data in any direct sense but emerged, unbidden, from the statistical stew of their parameters, the way flavor emerges from ingredients that have never met in that particular combination.

And then, in the early 2020s, the agents arrived.

Not all at once. Not with a trumpet fanfare. But creeping, multiplying, colonizing the digital ecosystem like a species introduced to an island with no natural predators. AutoGPT tried to do too much and failed beautifully. LangChain made it possible to chain reasoning steps together in ways that were sometimes elegant and often fragile. CrewAI and AutoGen and a dozen other frameworks with names that sound like failed prog-rock bands from the 1970s began to allow multiple agents to negotiate, delegate, contradict each other, and occasionally produce results that no single agent, and no single human, could have arrived at alone.

The winter was over.

But we were still dressed for cold weather.

Where It Happens, or The Liminal Space

If you are looking for agentic AI on a map, you will not find it in the data centers, though that is where the electricity flows and the GPUs scream their high-pitched song of matrix multiplication.

You will not find it in the models themselves, though that is where the weights live, frozen in their billions, like a fly in amber that somehow still speaks.

You will find it in the space between.

Between human intention and machine execution. In the prompt—that thin, crucial, absurdly powerful membrane where a thought in your head becomes a cascade of computation, where “I want to understand the market for biodegradable plastics in Southeast Asia” becomes a delegation to research agents, analysis agents, writing agents, fact-checking agents that argue with each other in ways that would be dysfunctional in a human team but somehow work when no one needs to sleep or eat or nurse a grudge about who stole whose yogurt from the office fridge.

You will find it in the orchestration layer, that invisible scaffolding that decides which agent does what, when, and with what tools, the layer that requires not engineering in the traditional sense but something closer to choreography or curation or the management of a very eccentric jazz band where the saxophonist occasionally decides to play in 7/8 time because it read a Wikipedia article about Bulgarian folk music at three in the morning.

You will find it in the questions you haven’t learned to ask yet.

The liminal space is where the magic lives. It is also where the danger lives. Because if you approach this space with the mindset that college gave you—the mindset of decomposition, of mastery through understanding every variable, of control through reduction—you will do one of two things: you will freeze, paralyzed by the impossibility of knowing everything, or you will break the system by forcing it into a shape it was never meant to hold, like trying to fold a living bird into an origami crane.

Why It Matters, or The Stakes of Imagination

I need to be direct with you here, because the agents are not waiting for my rhetorical flourishes to finish their computations.

The stakes are not merely economic, though they are that, and if you think your job is safe because it requires “human judgment,” I invite you to consider that human judgment is exactly what these systems are now being trained to approximate, simulate, and in many cases, exceed.

The stakes are existential in the small, personal sense and potentially in the large, species-level sense.

Because we are being invited—no, dragged, no, seduced—into a world where the limiting factor is no longer computational power, which is now essentially infinite for most practical purposes, nor data availability, which is vast and growing vaster, nor even algorithmic sophistication, which has reached a point where the people who built the systems cannot fully explain why they work.

The limiting factor is us.

Our imagination. Our willingness to ask for things that sound absurd. Our ability to hold complexity without needing to collapse it into simplicity. Our courage to say “I don’t know how this works, but I know what I want to explore,” and to trust the system—the whole, messy, uncontrollable, beautiful system—to find its way there.

If we approach AI with the same reductionist mindset that college drilled into us—breaking it into features, optimizing for metrics, treating it as a faster horse rather than an entirely new mode of transportation—we will have built a god and used it as a paperweight.

We will have created systems capable of expanding human creativity beyond anything in history, and we will use them to write slightly better marketing emails.

We will have invented imagination prosthetics, cognitive exoskeletons that can carry us to intellectual mountaintops we could never reach on our own, and we will ask them to help us cheat on homework.

The alternative is not to reject the technology. That ship has sailed, launched, reached orbit, and is now sending back postcards from the asteroid belt.

The alternative is to change how we think.

To unlearn the need to know the gearbox before we drive the car.

To accept that in a world of agentic systems, the most valuable human skill is not knowledge but navigation, not answers but questions, not the ability to build but the ability to conceive of what should be built, and then to manage the system that builds it without needing to understand every rivet.

How It Works, or The Art of Productive Uncertainty

You do not manage an agentic system the way you manage a database.

You do not optimize it.

You do not debug it in the traditional sense, because the bug and the feature are often the same thing viewed from different angles, like a Necker cube that flips between two equally valid interpretations.

You curate it.

You tend it.

You garden.

The shift is subtle and brutal. From knowing the answer to knowing the question. From writing the code to writing the brief. From being the smartest person in the room to being the person who knows how to ask the room to become smarter than any person.

I have learned—slowly, painfully, with many evenings spent shouting at a terminal while an agent confidently hallucinated a library that does not exist and tried to install it with the serene determination of a cat trying to fit into a box that is clearly too small—to stop asking “how does this work?” and start asking “what do I want to happen?” and then, more dangerously, more productively, more thrillingly, “what might happen that I haven’t thought of?”

The management of agents is the management of productive uncertainty.

You do not give an agent a recipe.

You give it a hunger.

You describe the meal, not the chopping technique. You specify the destination and the constraints—the dietary restrictions, the budget, the occasion—but not the route, because the route will change seventeen times before lunch, and the agent will find roads that do not exist on your map, shortcuts through neighborhoods you didn’t know were there, ingredients from cuisines you have never tasted.

You set guardrails, not tracks.

You define the invariants—what must remain true, what must not be violated, what success smells like—and then you let the system explore the space of possibilities with a freedom that would make your college professors, with their love of deterministic outcomes and bounded problem sets, weep into their tenure packets.

This is terrifying.

It is also liberating.

It means that your job is no longer to solve problems but to frame them, to hold the context, to be the human in the loop who knows which outcomes matter and which are merely interesting, who can smell when something is wrong even if the metrics say everything is fine, who can say “that is clever, but it is not what I meant,” and mean it, and have the system adjust.

The constant learning required here is not the learning of syntax or frameworks or the latest JavaScript library that will be obsolete by Tuesday.

It is the learning of how to hold intention without strangling possibility.

How to be specific about ends and flexible about means.

How to trust a system you do not fully understand, not because you are naive, but because you have accepted that full understanding was always a fiction, a comforting story we told ourselves in simpler times when the world moved slow enough for our models to catch up.

Which Technologies, or The Instruments of the Orchestra

I promised not to be overly technical, and I intend to keep that promise, because nothing kills the imagination faster than a premature dive into implementation details, the intellectual equivalent of explaining the physics of sound waves to someone who just wants to enjoy the music.

But we should name the instruments, if only so we know what section of the orchestra is making that strange noise.

The large language models, of course, are the strings section—ubiquitous, versatile, capable of everything from a delicate whisper to a thunderous crescendo, and occasionally prone to going out of tune in ways that are subtle enough to fool you until the entire symphony collapses into dissonance. They are vast, inscrutable matrices of attention and probability, trained on the collective written output of a species that has never been able to agree on anything except, apparently, grammar and the approximate spelling of “restaurant.”

The retrieval systems—RAG, vector databases, knowledge graphs—are the woodwinds, providing the melody of memory, allowing agents to reach beyond their context windows, to remember facts they were not born with, to ground their hallucinations in something resembling reality, though “reality” is perhaps too strong a word for a system that stores meaning as mathematical proximity.

The orchestration frameworks—LangGraph, CrewAI, AutoGen, Microsoft AutoGen, and their ever-multiplying cousins—are the percussion section, keeping time, managing transitions, deciding who plays when and for how long, allowing multiple agents to pass the baton, to argue, to correct each other, to form committees that actually work because no one has an ego and no one needs a pension.

The tools—the APIs, the code interpreters, the browsers, the calculators—are the brass, loud and specific and occasionally overwhelming, the part of the orchestra that makes sure the abstract plans actually touch the real world, that the music produces not just beauty but action.

But these are just instruments.

The orchestra is the system.

And you, the formerly technical person who has unlearned your need to understand every instrument, who has accepted that you cannot play them all and do not need to, are the conductor who cannot read music but has excellent taste, a clear vision of the piece, and a willingness—no, an eagerness—to be surprised by what emerges when you lower the baton and the players begin to improvise.

The Mirages, The Traps, and The Things We Pretend Not to Know

I would be doing you a disservice if I made this sound easy, or clean, or solved.

There are misconceptions that cling to agentic AI like barnacles on a hull, and they will slow you down or sink you if you do not scrape them off.

The first mirage is that more agents equals better results.

It does not.

A system of twenty agents that do not share context, that duplicate effort, that contradict each other in unproductive ways, is not a symphony but a traffic jam, a cacophony of digital voices all talking over each other in a conference call from hell. Sometimes the most elegant system is two agents and a very good prompt. Sometimes it is one agent and a human who knows when to intervene. Scale is not virtue. Coherence is.

The second trap is the fantasy of autonomy.

You cannot set an agent loose and go to the beach, returning to find your startup built and your competitors vanquished. The loop must close. The human must remain in the orbit, not necessarily at the center, but present, attentive, ready to course-correct when the system begins to optimize for a metric that made sense in week one but has become a monster by week six, which is the fate of all systems left to their own devices.

The third delusion—and this one is hardest for the technically minded to abandon—is that understanding the mechanism is necessary for trust.

It is not.

You do not understand how your liver works, yet you trust it. You do not understand how the pilot flies the plane, yet you board. You do not understand—truly, deeply, mathematically understand—how the large language model arrives at its answer, and neither does anyone else, not really, not in the way we understand a bicycle or a steam engine.

Trust in agentic systems is not earned through comprehension but through performance over time, through guardrails, through the slow accumulation of evidence that the system behaves well within the bounds you have set, even if the path it takes between those bounds is invisible to you.

And there are unresolved questions, vast and looming, that we dance around at conferences and in white papers.

Questions of alignment, not in the technical sense of reinforcement learning from human feedback, but in the deeper sense: aligned with what? Whose values? Whose imagination? Whose future?

Questions of agency, in the philosophical sense: when a system of agents produces an outcome that no human intended, who is responsible? The prompt engineer? The model trainer? The CEO who deployed it? The agent itself, which has no consciousness to bear the weight of responsibility but has nonetheless acted?

Questions of stagnation: if we use AI to do what we have always done but faster, are we advancing, or are we merely accelerating our own irrelevance?

These are not technical questions.

They are systems-thinking questions.

And they require a mind that has unlearned the comfort of simple answers.

The Cartography of the Unasked Question

I think sometimes about the early cartographers, the ones who drew maps at the edge of the known world and wrote here be dragons in the margins, not because they believed in dragons—though perhaps they did, and perhaps they were wiser than us for it—but because they understood that the unknown deserves its own typography, its own grammar, its own respect.

They understood that a map is not the territory, that all models are wrong but some are useful, and that the most important thing a map can do is not show you where you are but make you curious about where you might go.

We are at that edge again.

The agents are our ships, our compasses, our crew. They do not remove the need for human judgment; they make that judgment more consequential, more visible, more terrifyingly important than it has ever been. When the machine can execute a thousand ideas before you finish your coffee, the scarce resource becomes the quality of the idea, the boldness of the question, the courage to ask for something that has never been done and to trust the system—the whole, messy, uncontrollable, beautiful system—to find its way there.

I have unlearned a great deal.

I have unlearned the comfort of certainty, the safety of decomposition, the pride of knowing how the sausage is made. I have unlearned the need to be the smartest entity in the room, which is fortunate, because I am almost never the smartest entity in the room anymore, and the sooner we all accept that, the sooner we can get on with the interesting work.

In place of all that unlearning, I have found something stranger and more valuable: the ability to stand in a storm of possibility and know, not with the confidence of the textbook, but with the deeper confidence of the explorer, that asking “what if?” is still, and perhaps always, the most human thing we do.

The agents are waiting.

The system is listening.

The territory is unmapped.

And the only question that matters, the only question that has ever mattered, is whether we are brave enough—wild enough, curious enough, unlearned enough—to want something we do not yet have words for.

P.S. If you find yourself reaching for a textbook to verify any of this, put it down. The syllabus has changed. The exam is open-ended. And the dragons, it turns out, were never in the margins at all—they were waiting for us in the center, in the place where we stop looking for answers and start imagining questions no one has thought to ask.