When I cofounded Concentric AI, the first guiding principle I wrote down<\/a> for myself was first-principles thinking. Ignore the conventional wisdom about what is possible. Understand the real problem. Follow it wherever it leads.<\/p>\n\n\n\n I came back to that method for the book. When I introduced this newsletter, I put it plainly<\/a>. There is only one way I know to cut through noise. You stop arguing at the surface and go back to first principles. You find the single thing that actually changed, and you follow it, patiently, wherever it leads, whether or not the destination is fashionable.<\/p>\n\n\n\n The single thing that has changed now is the cost of execution. It has collapsed toward zero. So I started there and followed it, step by step. The first step was a question that sounds too basic to be useful. Why do firms exist at all?<\/p>\n\n\n\n Someone answered that in 1937.<\/p>\n\n\n\n A twenty-six-year-old economist named Ronald Coase<\/a> published a short paper called “The Nature of the Firm<\/a>.” He had been carrying the idea since his undergraduate years, when he toured American factories trying to work out why industry was organized the way it was. He later called the paper “little more than an undergraduate essay<\/a>.” That essay was a big part of why he was awarded the Nobel Prize in 1991. It is one of the cleanest pieces of first-principles thinking I have come across.<\/p>\n\n\n\n Coase asked a question no one had bothered to ask. Economics of the day said markets coordinate activity efficiently. In wikipedia’s words, that means those who are best at providing each good or service most cheaply are already doing so. If that were the whole story, there would be no need for companies. Every task would be handled by independent people contracting with each other in the open market. So why do firms exist at all, with their managers and hierarchies and payrolls?<\/p>\n\n\n\n His answer was transaction costs. Using the market is not free. You have to find the right person, agree a price, write the contract, monitor the work, enforce the terms. When doing all of that inside a company costs less than doing it through the market, the company does it inside. Firms exist because internal coordination is sometimes cheaper than market coordination. The boundary of a firm sits exactly where those two costs meet.<\/p>\n\n\n\n That was the insight. A firm is an efficiency solution to the cost of getting things done. And it carried a quiet implication that took decades to surface. When the cost of coordination changes, the best shape for the organization changes with it.<\/p>\n\n\n\n Which brings us to now.<\/p>\n\n\n\n For nearly a century, the binding constraint inside a firm was execution. Doing the work took scarce, expensive people. Building software, processing information, running an analysis, deploying a system, all of it needed specialists. Much of the org chart existed to allocate that scarce execution. Management decided what got built, by whom, in what order.<\/p>\n\n\n\n Agentic AI removes that constraint. Execution is collapsing toward free. A company’s ability to build workflows, deploy agents, and automate complex processes is no longer capped by engineering capacity. The scarcity the hierarchy was built to manage is dissolving.<\/p>\n\n\n\n What is left, and what now binds, is coordination. Keeping abundant autonomous systems coherent, steerable, and pointed at one purpose is getting harder and more expensive, in exactly the ways existing structures were never designed to handle.<\/p>\n\n\n\n This is what the book calls the Coasian Inversion<\/a>. The firm was built to economize on scarce execution. Agentic AI breaks the assumptions underneath that design. Execution, the thing firms were built to marshal, is now cheap. Coordination, the thing they were never built to manage, is now the constraint.<\/p>\n\n\n\n The AI era does not end the need for firms. It changes the reason they exist. They no longer exist mainly to allocate scarce execution. They exist to hold abundant autonomous capability together. That is a different job, and the structures built for the old one are not built for the new one.<\/p>\n\n\n\n That is where following one changed cost leads. The answer is structural, and it began for me with an undergraduate essay.<\/p>\n\n\n\n Three posts from the past week.<\/p>\n\n\n\n I studied the Brain to Build AI. The Agentic Era Sent Me Back to It.<\/strong><\/a> I trained in neuroscience before AI, and it left me a habit of asking what a technology can really do. I was skeptical of agentic AI, because reliability does not survive being chained. Stack ten agents at ninety-five percent each and you end up near sixty. Then coding agents improved so fast that I changed my mind about the timing. Once you can trust individual agents, you do not stop at one. You deploy fleets, and a new question appears that has nothing to do with reliability. If every agent does its job, does the fleet still add up to what the organization intended? A body answers that with proprioception, the constant inner sense of where its parts are. An organization needs the same inner sense, or capable parts never combine into coordinated action. Weigh in on LinkedIn\u2026<\/a><\/p>\n\n\n\n A Company Doesn’t Have One Brain.<\/strong><\/a> The conversation has moved fast. “Buy the best model” is giving way to “protect your proprietary layer,” what BCG now calls the enterprise cortex. I agree that the model is a commodity and the moat is what the organization knows about itself. But the cortex framing hides an assumption. A brain has one cortex. An enterprise grows a dozen, each team with its own definitions and rules, and no one owns how they combine. You can own every byte and keep every vendor out, and still fail, because your pricing logic and your inventory logic never agreed on what a lapsed customer is. Owning your brain and organizing your brain are two different jobs. Knowing is not the same as coherising. Weigh in on LinkedIn\u2026<\/a><\/p>\n\n\n\n What the Watermark Doesn’t Tell You.<\/strong><\/a> Anthropic began watermarking the text Claude writes, and detectors are spreading. They all read one thing, which tool produced the words. None of them reads whether the writing is any good. Slop is not a provenance problem. It is a failure of intent or execution, and no watermark reads either one. A person can produce slop unaided. Someone with a clear point can use AI to sharpen it and produce something worth reading. Tracing the tool cannot tell the two apart. Weigh in on LinkedIn\u2026<\/a><\/p>\n\n\n\n One thread ties these to the essay above. Each is a dispatch from the new constraint. The brain piece names what an organization needs to stay coordinated, an inner sense of whether its parts still add up to intent. The cortex piece shows that owning the parts is not the same as making them agree. The watermark piece shows that no external mark can tell you whether the parts hold together, because that is a question of judgment, not provenance. Cheap execution handed every company more parts than it can see. Keeping them coherent is the work now.<\/p>\n\n\n\nThe week in ideas<\/h2>\n\n\n\n
Before you go<\/h2>\n\n\n\n