I’ve been building my company a certain way for a while without a name for it. Then I found a book that named it, and now I can’t unsee the pattern. I want to put the idea down while I’m still in the middle of testing it, because I’m not certain it holds, and writing it down is part of how I find out.
Here is the concept. Most of us deploy AI like a machine. I think the better model is an octopus. I’ll tell you what I mean and why I’m betting on it, and I’ll be honest about where I might be wrong.
The machine version is almost everywhere
Start with the machine version, because it’s the default. One big workflow, mapped end to end before anyone starts. A central agent that every request routes through. A standardized prompt that each task has to fit. The whole thing built to run the same way every time, with as little variation as possible. It feels right because it looks like a machine, and a machine is what we know how to build.
I built things this way for years. It’s a hard habit to see from the inside, which is part of why I think it’s worth naming.
What the octopus does differently
Phil Le-Brun and Jana Werner, both enterprise strategists at AWS, name the alternative in The Octopus Organization (Harvard Business Review Press, 2025) and the companion article Become an octopus organization. They’re writing about organizations, not AI specifically, but the moment I discovered it I saw my own setup in it.
An octopus keeps about two-thirds of its neurons in its arms, not its head. Each arm senses and acts on its own slice of the world without waiting for central command. The brain sets intent and handles the genuinely novel problems, the ones an arm can’t resolve on its own. Coordination doesn’t run through constant top-down messaging. It runs through a shared surface, the skin, which reacts to local conditions and still produces one coherent body.
What it looks like in practice
Here is what the octopus model looks like in practice, the way I’m working with it. Each role I hand to AI gets its own slice: its own job, its own memory, its own tools, its own feedback loop. I set the intent and take the calls the roles can’t. They don’t talk to each other all day. They read from and write to a shared context surface, the equivalent of the octopus skin: the logs, the memory files, the playbooks. I keep that surface current, and the rest mostly coordinates itself.
What I keep noticing is that the machine instinct is the thing that breaks it. Every time I over-specify a step, I kill the judgment I wanted. Every time I route a decision back through myself, I become the bottleneck the whole design was supposed to remove.
This is the same thing I was reaching for when I wrote that AI needs a job description, not a better prompt. A role that owns a responsibility holds quality across many runs. A task is a cold start every time. The octopus frame just tells me where to put the roles: at the edges, with real autonomy, not stacked underneath a central controller.
The anti-patterns I catch myself in
The book names thirty-six anti-patterns, deep habits that quietly cost a company its clarity, ownership, and curiosity. A few map straight onto AI, and I catch myself in them more than I’d like.
Scripting every prompt step. Designing a giant multi-agent workflow up front, before a single role has earned its place. Standardizing the output when the value was in the context-sensitive adaptation. Routing every call back through one agent, or through me. They all come from the same reflex: reach for control, and call it rigor.
Where the value moves
The implication I keep sitting with is about where the value goes. In a machine organization, the talent is the arms. People execute, management coordinates. In an octopus AI organization, the talent is the brain.
The execution moves to the agents. What’s left for the humans is the brain work: setting intent, designing the experiments, reading the patterns, tending the shared fabric that keeps the whole thing coherent. I don’t fully know yet what that means for how I hire or how I spend my own day. I just notice my own time bending that way, away from doing the arm work and toward designing the system that does it.
It rhymes with something I already believe, that marketing is infrastructure, not headcount. The octopus frame is that same bet one layer down: the next decade gets measured in systems built, not bodies hired, and now not in agents bolted on either. Bolt a fleet of agents onto a machine-shaped operation and I think you just multiply the breakage faster. Though I hold that one loosely too.
Where I might be wrong
Here is where I have to be straight about the edges of this. Le-Brun and Werner wrote for Fortune 500 enterprises. I’m applying their frame to a small operation where most of the workers are literally AI agents, which is further than they took it. The scaffold feels right. The extension is mine, and it’s unproven.
I’m running my own company on this bet right now, which means I’ll find out whether it holds the way you find out anything: by living inside it and watching what breaks. So treat this as a position I’m testing, not a conclusion I’m selling. I might read it back in a year and wince. If I do, I’ll say so. The whole point of the octopus is adapting to what the water is actually doing, and that has to include the idea itself.
Who this isn’t for
If what you want is one big automation that runs the business while you step away from it entirely, this probably reads as more moving parts than you hoped for. I understand the appeal. I just don’t believe it anymore. The black box that does everything and explains nothing is the machine model in its purest form, and the machine model is the thing I watched stop working.
I’m not certain about most of this. I’m certain enough to build on it, and honest enough to tell you it’s a bet. Ask me in a year. I’ll have the data by then, because I’m the one living in the experiment.
