The comfortable shape is a triangle.
Food pyramid. Maslow’s hierarchy. The org chart on slide two. CEO at the apex, layers beneath, maybe an arrow labeled “bottom-up” if leadership read the memo that top-down bossing is out of fashion. Matthew Barzun’s opening move on Invisible Machines is to treat that shape as the enemy of the story—not because order is bad, but because we’ve mistaken one kind of order for the only kind.
Maslow never drew a triangle. Late in life he said needs overlap; they don’t stack. Kentucky doesn’t report to Washington. Berkeley doesn’t report to Sacramento. At every scale, Barzun argues, a different pattern already works: the constellation—you stand out as a star, you connect to other stars, you become something bigger than any single node. The United States, in his framing, is a fractal history of that pattern. So is the internet. So is any marriage that survives more than a quarter.
This matters now because agentic AI is being discussed almost exclusively in pyramid terms: Will these systems be our overlords or our tools? Who’s on top? Who reports to whom? Barzun’s answer, developed in The Power of Giving Away Power and sharpened across two conversations with Josh Tyson and Robb Wilson, is that both choices are the same mistake—a triangle wearing different makeup.
The Pyramid Trap
When the CEO gets tired of lording power, the reflex is to flip the triangle—bottom-up reviews, inverted hierarchies, “empowerment” theater. Barzun’s line is blunt: you’re still asking every person on the team to locate themselves above or below someone else. Same shape. Same mental model. The connectivity that actually moves work—who do you need to talk to, what are you trying to do—never enters the room.
Robb Wilson names the design parallel immediately. In agentic systems we obsess over instructions because that’s how software used to work: specify the steps, guarantee the behavior. LLMs invert the problem—you feed objectives, not procedures, and the system invents paths you didn’t document. Organizations still importing pyramid habits into that stack will build agents that rank, sort, and comply when the opportunity is to connect, integrate, and re-solve.
Barzun barely uses the word power in daily life; the book title uses it twice. His taxonomy is clearer: you can lord power over people, hoard it, or divvy it up in well-intentioned fairness—which still assumes power is finite, like coal in Kentucky. The leaders he admires treat power as infinite when made with and through others. That isn’t a bumper sticker. It’s a design stance for agentic AI: the tool isn’t your boss or your servant; it’s a participant in co-creation where the human still owns the outcome they care about.
Mary Parker Follett and Integration
Mary Parker Follett—Peter Drucker’s guru, erased from the guru lists—argued a century ago against scientific management measuring factory strides while treating people as machines. Her meeting story is the one to steal for product teams. The electrician says Mary’s wiring plan will burn the house down. He proposes an alternative she refuses—I won’t live in a house lit like that; it’s unlivable. Compromise leaves half the house burned and half unlivable. Integration is the third thing: a plan that meets code and feels livable, invented in the encounter, not carried in by either party.
Josh Tyson connects it to what individuals are already doing with AI while enterprises lag: co-creating workflows, agents, and personal systems—not compromising with a chatbot, making something with it. Barzun extends the thread to journalism at The Observer: use a model to draft what well-intentioned journalists get wrong about an industry, then sit with practitioners and integrate—machine-assisted beginning, human integration, no machine in the room for the part that counts.
Fight the Mechanical
Barzun’s grandfather Jacques Barzun devoted his life to fight the mechanical—not machines, but the mindset that ranks, sorts, and automates the good parts of being human. The episode stacks pyramid myths that became culture: Darwin’s “survival of the fittest,” the alpha wolf paper based on captive animals. Follett fought Frederick Taylor’s stopwatch; today we fight outsourcing judgment and calling it progress.
Donella Meadows supplies the systems vocabulary: we grab the obvious knob and wonder why the bathtub stays cold. Higher leverage lives in rules, goals, and mindsets. Barzun hears Robb and Josh on hierarchy and adds: the cruelest pyramids are illusions. People think there’s a ladder; there isn’t; everyone brings a hobbled subset of self to fit a rung that doesn’t exist.
Constellation Proof Cases
Dee Hock built Visa as tiny banks in a constellation, not a supernode. Vint Cerf, visiting the London embassy after Snowden, told young activists the internet was a verb—Would you like to internet with me?—and TCP/IP became the most reliable network ever by letting connections go. That’s the design lesson for agentic stacks: reliability through release, not grip.
Southwest Airlines, in a snowstorm, externalizes the constraint on a clipboard—Louisville canceled, Lexington canceled—and stands shoulder to shoulder with you: We’ve got to get you home. On Zoom, everyone embodies the problem in their square. Whiteboards and AI drafts disembody it—something that isn’t you, that you can edit together.
Constructive Tension
Co-creation sounds like a group hug; Barzun offers re-solve instead of solve. His 2×2 names the failure mode: fight it out, hug it out, sit it out—the triangle of sadness. The exit is constructive tension: argue without trying to win; delight in difference. That’s why always-agreeable AI irritates—it’s pandering death.
Ken Burns said he tells his team: never sound like a math teacher. Barzun used clickers with 20,000 UK students so the ambassador couldn’t lecture. Agentic AI doesn’t require a new politics. It requires admitting the politics we keep drawing is wrong.
The design brief isn’t “who’s the boss of the agents.” It’s whether your organization can surface difference safely, externalize constraints, co-create objectives with the people already playing at the edges, and treat every “solution” as a re-solve—good for now, wrong later, ready for the next pass.
Open the Ideation hub, or read the full episode transcript.