Dismantling Silos: A Path to Agile Engineering

Boundaries are necessary. That’s not the argument.

Every engineering project starts with bounding — what you’re solving, what the solution has to do, what’s out of scope. Without that, you’re not engineering, you’re wandering. The boundary is how you make the problem solvable.

The modern corporation learned the same lesson at scale. Adam Smith’s insight wasn’t complicated: split work into elements, run them in parallel, and you can deliver what no individual craftsman ever could. From Renaissance capital markets to the factory floor to the aerospace prime contractor, that logic held. Boundaries enabled scale.

When I joined the workforce in 1982, the logic was still holding — and you could feel why. I had a notebook and an HP calculator. A shared secretary supported the division manager, and before any report left the building it needed sign-off from both my branch manager and his. Not bureaucratic obstruction — that was the information architecture. Reports were dense, slow, and gatekept because they had to be. Management structure existed in large part to curate that flow — to compress what mattered, pass it up the chain, and keep the organization pointed in the right direction. The stovepipe wasn’t a bug. It was load-bearing.

Between 1982 and 2002 two things happened simultaneously that should have changed the equation. First, information handling exploded. The PC, networks, sensors — generating and moving information became cheap and fast. Second, process culture arrived. The US had watched the Japanese manufacturing renaissance and brought back a set of ideas about quality and process that got bolted onto the existing corporate hierarchy. At exactly the moment when individual engineers could span across an organization and get at information directly, the process culture locked the structure down harder.

The result in many companies: more capability to move information, less permission to use it. The stovepipes stayed. The rationale quietly expired.

I ran three programs across my career that show the delta. At SatCon on the AIPM program — Advanced Integrated Power Module, a DOE/Navy cost-share — I was simultaneously program manager and lead engineer, spanning manufacturing, electrical design, mechanical design, and simulation. We went from concept to demonstrated production-ready modules in three years on a modest budget. That approach, the sub-module test-before-integrate architecture we developed, is now standard inside automotive power electronics. Tesla uses a version of it.

At DRS, working with Allison Transmission on an integrated generator for military vehicles, we built a successful solution and demonstrated it to the Army. General officers asked why they couldn’t have more. It took ten years for the technology to gain traction — not because the engineering was wrong, but because the organizational and procurement structure couldn’t move.

At Wolfspeed, deep stovepipes. Marketing, sales, test engineering, module design, device fabrication — separate organizations, separate priorities, separate permission structures. Getting a new product from concept to release meant handing information off at each boundary and then jawboning it forward, because you couldn’t do their job for them and they had to queue the work against their own priorities. Fifteen products out the door. Every one of them harder than it needed to be.

The stovepipes were there to protect quality. They also stopped momentum.

What’s changed now isn’t the human desire to span boundaries — engineers have always wanted to do that. What’s changed is that the tools exist to actually do it. Companies that have built their information architecture from scratch rather than inheriting it — the Teslas, the newer defense tech firms — have demonstrated what happens when low-level actors have access to the full context of what the organization knows. Engineers and technicians can interrogate data, surface patterns, propose action. The information that used to require a management layer to curate is available directly. The span of control moves down the org chart.

For incumbent organizations with data already siloed, this is genuinely hard. The stovepipes aren’t just structural — they’re also where the institutional knowledge lives, and dismantling them requires executives who are willing to accept that the curation function they’ve been performing can be partially replaced. That’s not a technical problem. It’s a political one.

Christensen’s Innovator’s Dilemma describes what happens to incumbents who don’t solve it. A smaller firm with narrower scope but faster movement finds a niche. The niche gets cheaper and easier to serve. The incumbent can’t see it clearly because their whole architecture is optimized for something else. The niche expands. You know the rest.

The boundary isn’t the problem. Bounding a problem is still part of the engineering job. The question is whether, once the problem is bounded and the work begins, you’re working inside a structure that moves — or one that fills up and waits to overflow into the next pipe.

While many organizations are ‘implementing AI’ most are not working through the changes from first principles and often implementing all or nothing. The ones that don’t get around to making sure they break the stovepipes logically are going to run out of time.


This post accompanies the video Why Stovepipe Organizations Stop Working — The Unretired Engineer, April 2026.

How Physics Empowers Free Will in a Deterministic Universe

Why determinism never felt right to me — and how modern physics actually opens the door to real agency.

For years, the idea of hard determinism has bothered me. It clashes with how life actually feels. The universe as a giant clockwork machine—every particle with a fixed position and momentum, every event preordained from the Big Bang—sounds elegant in theory. But it implies that everything I’ve ever done or will do was inevitable. My choices? Just an illusion.

Hard determinists often present this view with a certain intellectual swagger, as if it signals deep sophistication. Yet many of them still look both ways before crossing the street. As Stephen Hawking wryly observed: “I have noticed that even those who assert that everything is predestined… still look both ways before they cross the road.”

That quip captures the tension. If the future is fixed, why bother acting at all? The view also carries an eerie resemblance to extreme Calvinism—some are saved, some damned, and nothing you do in this life ultimately changes the script. It never sat right with me, either intellectually or existentially.

Then I encountered the work of physicist and philosopher Jenann Ismael, particularly her book How Physics Makes Us Free. Link Her approach resonated strongly with an intuition I’d been developing for years: determinism and free will are not mutually exclusive. Physics doesn’t enslave us—it enables a deeper kind of freedom.

The “Now” Problem: Why the Instant is Trivial

Imagine the universe at a single frozen instant—the “Now.” In that timeless 3D snapshot, every particle has a position and energy vector. Past events fully determine what happens next. It looks perfectly deterministic.

But here’s the catch: that “Now” has no real existence for any actual observer. Relativity imposes strict limits. No particle (or person) can access information from outside its past light cone. At the exact moment of “Now,” that light cone has zero depth—nothing from even a tiny distance away has had time to reach you. Complete information about the universe is impossible in the present.

Laplace’s Demon—the hypothetical super-intellect that knows every particle’s state and can predict the entire future (or past)—assumes a “God’s-eye view from nowhere.” Modern physics makes that view untenable. Any real system faces data latency, noise, uncertainty, and computational irreducibility. The demon’s omniscience is a fantasy.

In short, strict determinism at the instantaneous “Now” (what I’ve called the InP, or Instant-Point) is technically true but functionally trivial. It tells us almost nothing useful about how agents like us actually operate.

Memory: The Engine of Agency

Freedom emerges not in the frozen instant, but across time through accumulated memory and structure.

Even in a blind, non-living universe, basic thermodynamics creates imprints: a rock scars the ground when it falls; waves erode a shoreline. These are primitive forms of “memory”—the past shaping the future through persistent physical traces.

Life takes this to another level. Biology is essentially memory in action. RNA, DNA, neural patterns—these are systems that record what worked and what didn’t. Evolution itself is a memory process: successful patterns persist and build upon one another.

Over eons, this scales up:

  • Simple input → output (basic matter)
  • Input → memory/comparison → internal model → action → output (living organisms)

A frog snaps at a fly. A squirrel flees at a predator’s scent. A honeybee dances to communicate nectar locations to the hive. These are not random reflexes but decisions grounded in accumulated history and pattern-matching.

Humans take it further. Language, culture, and shared knowledge externalize memory, allowing us to build on the experiences of countless others. Our decisions arise from rich internal models shaped by personal and collective history—not from some magical spark that violates physics, but from the universe’s own lawful processes.

The agent does decide. The cause of the action lies in the person’s internal identity and accumulated experience. Labeling that “determined” is technically accurate but misses the point. It’s how we function.

The Generalized Good as an Attractor

This memory-driven agency isn’t aimless. Over deep time, beings with even modest volition tend to optimize for what they perceive as “good”—survival, order, flourishing. Humans are guidable, not perfectible. We make mistakes and fall for bad influences, but signals from reality (what works vs. what fails catastrophically) are powerful if we’re willing to heed them.

History shows progress: fewer people in extreme poverty, fewer dying in wars (in percentage terms, at least). Our ancestors weren’t ignorant fools; their traditions often encoded hard-won lessons. Change isn’t inherently good, but neither is stasis. The “generalized good” acts as a global attractor, even if local maxima vary by time, place, and culture.

In deprived environments (think North Korea), external options shrink, yet people still imagine and yearn for “other worlds.” The internal model remains a generator of possibility.

My Thesis

Free will is not a violation of physics. It is the high-level, computational process of an autonomous agent using the universe’s built-in memory—personal, biological, and cultural—to steer itself through time.

Determinism at the microscopic level may hold, but it becomes trivial once you account for relativity, light cones, computational limits, and the reality of embedded agents. What matters is that you are the one deciding, drawing on your history and internal model. There is no external puppet master. The causes flow through you.

Physics doesn’t rob us of freedom. By creating a world with persistent memory, evolving complexity, and embedded perspectives, it makes genuine agency possible.

That’s why the universe feels open rather than claustrophobic. That’s how physics makes us free.

The Problem With AI Answers Is That They’re Almost Right

AI slop isn’t obvious. That’s what makes it dangerous.

If an AI gave you complete nonsense, you’d catch it. The problem is when it gives you something fluent, confident, and “mostly” correct — with a flaw buried in the middle that you’ll only find if you already know the answer.

That’s the thing about AI as a research tool: it will give you the consensus view, coherently expressed, at the level of resolution that the training data supports. Where the training data is thin, ambiguous, or where real expertise requires distinguishing between things that *look* similar but aren’t — that’s where it fails. And it fails confidently.

Even when you use the deep research tools there are problems. When I was developing some content for my YouTube channel, The Unretired Engineer I ran into this doing research on Wolfspeed’s financial situation and the SiC power electronics market. I asked a deep research tool to pull together an analysis. What came back looked thorough. The problem was that it took a lot of information that had gone out about the future of the fab and future plans for markets and conflated them with what had happened and what was likely to happen in the near future.

To someone without a background with Wolfspeed and the real status of the SiC, the analysis would have read as authoritative. It wasn’t. It had serious timing errors delivered with confidence. I knew it was wrong because I’d spent years in that space. If I hadn’t, I might have taken it as written.

The fix isn’t to stop using it. The fix is to put yourself into it.

When I work with AI on my engineering writing, or on the physics underlying my novels, I’m not asking it to do the thinking. I’m using my domain knowledge to steer it, to catch the near-misses, and to push it past the consensus into territory where the expertise actually matters. The AI amplifies what I bring. Without that, it’s just averaging.

Use it as a tool. But know what it can’t know — and that’s usually the thing that matters most.



https://youtube.com/shorts/mbmKm_JcHQ0?feature=share

Mark Harris is a system and mechanical engineer and the author of “Stranded in the Stars” (Book One, The Sea of Suns Trilogy), available now on [Amazon](https://www.amazon.com/Stranded-Stars-M-Harris-ebook/dp/B0GT123PLP)