Systems Thinking: The Engineering Skill AI Can't Replace

I work with AI every single day, and it has made me faster than I have ever been. It removes the labor: the boilerplate, the lookups, the first draft of almost everything. But the more it does, the more one uncomfortable question keeps circling back to me. If the AI does the work, what exactly is my job now?

I have sat with that question for a while, and I think the honest answer is also the reassuring one. What is left when the labor is gone is the part that was always the actual engineering. Not the typing. The thinking about the whole. And it has a name: systems thinking.

AI took the labor

Let me be clear about how much it really does, because pretending otherwise is silly. A huge share of what used to fill my days, scaffolding, glue code, looking up the exact syntax, writing the obvious first version, is now handled in seconds. We are all more productive because the manpower part of the job, the sheer volume of execution, has been lifted off our shoulders.

That is genuinely great. But productivity is not the same thing as value, and this is where it gets interesting. When the execution becomes cheap and fast for everyone, it stops being where you stand out. If everyone can produce the code, the question quietly shifts from "can you build it" to "do you know what should be built, and why, and how it all fits together."

I pushed this further by wiring up an AI agent with n8n that works through the bug backlog on its own. It picks up tickets, fixes them, and clears them while I am busy with something else, so the workload is lifted off in a very literal sense. But the part I care about is not the speed, it is what comes back: for every fix the agent writes up a report of what was wrong, how it was solved, and why. That paper trail is what keeps me in control. The execution is delegated, but the understanding is not, I still get the full picture of every change and can see how each one fits into the rest of the system.

So what is left?

Here is the part I find oddly comforting. The big tech companies are not, deep down, hiring for typing speed, and they never really were. What they look for in an engineer is someone who understands complex systems, who can take them apart, analyze, and see how one change ripples through everything else. Someone who can orchestrate, hold the whole picture in their head and make the parts move together toward a goal.

That is precisely what an AI cannot do. A model is a brilliant executor with no map of the territory. It will write you a flawless function for the wrong problem, with total confidence. It does not know why the work matters, what it connects to, or what breaks three layers down when you change this one thing. Holding that map is the human job. That is systems thinking.

The cognitive skill of the 21st century

Edward Crawley, a professor at MIT, called systems thinking the cognitive skill of the 21st century. I brushed past the line the first time I read it. Then AI moved into my daily work, and it stopped being abstract. It is exactly the skill this new way of working rewards, and I can feel it in how my days have shifted.

Crawley defines a system simply: the parts, and the relationships between them. That second half is the whole trap. AI is astonishingly good at the parts, it will hand me any single piece I ask for, fast and clean. What it does not hold is the relationships: how this piece pushes on that one, what emerges when they all run together, what quietly breaks three layers down. He describes real complexity as "more parts talking to more parts," and that web of conversations is exactly the part that never left my desk.

So my job quietly rewired itself, and this is the part I did not see coming. The work moved from producing the parts to reasoning about how they connect, from breaking a problem down and building each bit, to holding the whole in my head and watching the interactions. Crawley says learning to think this way grows "new neural pathways," and honestly, that is how it feels: I think about my own work differently than I did two years ago. Less "write the thing," more "see the system the thing lives in." That shift, not the speed, is what AI actually gave me.

How I actually work with AI

This is not theory for me, it is how I run my days, so let me show you concretely. I do not prompt randomly into a chat box and hope, like this:

That is a wish, not a system. I treat my AI setup as a system I have designed instead, and I think that is the whole point of this article in miniature.

First, everything is organized by folders, because structure is intent. The way I lay things out tells the AI, and me, what belongs together and what does not. Nothing is in one big pile waiting to confuse the model.

At the center of it I keep a dedicated agents folder, split by need. Instead of one generic assistant that is mediocre at everything, I have a set of specialized agents, each one carved out for a specific kind of task. One understands a particular domain, another reviews, another digs through data, and so on. I add a new one when a real, recurring need appears, not before.

And each of those agents is not improvised. Every one is built on actual research, scoped instructions, carefully chosen context, examples, so that it frames and maximizes its retrieval instead of drowning in irrelevant input. That is the real craft behind good RAG: not feeding the model everything, but giving it exactly the right knowledge for its job and deliberately keeping the noise out. A well-scoped agent with the right context beats a giant prompt with everything thrown in, every time.

On top of that, I have built my own modes to shape how the model behaves, rather than accepting its defaults. I run a caveman mode that strips its replies down to dense, no-filler signal, it saves tokens and, honestly, my reading time. A ponytail mode that forces the laziest solution that actually works and refuses to over-engineer, no speculative abstractions, no scaffolding "for later." And a memory layer, claude mem, so the AI remembers my preferences and the decisions we have already made instead of making me re-explain myself every session. None of these are features I was handed. They are knobs I added because I decided how I wanted the system to behave.

Then I orchestrate. I decide which agent handles which step, in what order, how their outputs feed each other, and where I need to step in and verify. I am not the laborer anymore. I am the conductor. And designing that little orchestra, deciding what each part does and how they connect, is itself an act of systems thinking applied to my own workflow.

The skill to actually build

So if I had one piece of advice for engineers right now, it would not be "learn to prompt faster." It would be: build your systems thinking. Get good at seeing wholes, at understanding how things connect, at deciding what matters and orchestrating the rest. Prompting, agents, RAG, all of it is just the instrument. The musician is the one who understands the whole piece.

AI raised the ceiling on what a single person can do. It did not remove the need for someone to decide what is worth doing and to hold the system together while it gets done. That someone is the engineer who thinks in systems. That is the skill that is becoming mandatory, and the one no model is going to take from us any time soon.