Coordination Is All You Need

For most of the last decade, getting more intelligence out of AI meant one thing: make the model bigger. More parameters, more data, more compute. Intelligence was something you trained into the weights, and the recipe for more of it was to scale. That recipe still works — but it has quietly stopped being where the hard, interesting problems live.

Intelligence has started leaking out of the weights and into the structure around them. The systems doing the most useful work today are not single models answering a prompt; they are arrangements — chains of model calls, tools, retrieved memories, and other agents, composed so the whole behaves more capably than any part. What each piece sees, what it remembers, who it hands off to and when: that wiring now carries as much of the system's intelligence as the weights do. The model has become a component. The architecture is the system. Intelligence is no longer only in the tokens and the weights — it is in the structure.

That is where the title comes from. The mechanism that made a single model intelligent was, in the end, just a way of routing information between its parts so each could attend to the others that mattered — intelligence inside a model turned out to be a coordination problem between its pieces. I have come to believe the same sentence describes the next layer up. The hard problem is no longer intelligence inside a model. It is coordination between models — and between the models and us.

I did not arrive at that by theory. I arrived at it by failing the same way, over and over.

How I found the bottleneck

For the last few years I built agents — browser agents, deployment agents, function-calling primitives, copilots. Each one demoed beautifully and then stalled at the same wall the moment I asked it to do sustained, real work. The stall was never a reasoning failure. The model was plenty smart. What it lacked was everything around the task: what had been decided last week, which document was the source of truth, who else was acting right now, what the goal actually was, and who needed to know when it changed. Drop a brilliant new hire into a company with no onboarding, no colleagues, and no memory of yesterday, and they will also look useless. Not because they are dim — because they are alone.

The first time I really saw it, I had stopped thinking of it as a prompt problem and started recognizing it as something older. An economist named Ronald Coase asked, back in 1937, a question that sounds almost childish: if markets are so efficient, why do companies exist at all? Why isn't every task just a contract between individuals? His answer, in The Nature of the Firm, was that using the market is not free. Every transaction carries a coordination cost — finding the right party, negotiating, establishing trust, transferring context. Firms exist because, past a certain point, it is cheaper to coordinate work inside an organization than to renegotiate it from scratch every time. The org chart is not bureaucracy. It is a machine for driving coordination cost down.

That was the click. A lone agent pays Coase's transaction cost on every single step — it reconstructs the world from nothing each session. An organization amortizes that cost across time and across members. The thing my agents were missing wasn't a bigger brain. It was a firm.

The bottleneck really did move

It is worth being precise about why capability stopped being scarce. The frontier labs all ship models within a few points of each other; you rent the same weights your competitor does. Meanwhile the failures that actually break agent systems are not failures of raw intelligence.

Two findings convinced me this is structural, not anecdotal. First, on the single-agent side: the obvious fix for "the agent forgot" is a bigger context window — just give it everything. But Lost in the Middle (Liu et al., 2023) showed that models use long contexts badly: accuracy is high when the relevant fact sits at the very start or end and sags when it is buried in the middle, and it degrades as the context grows — even for models explicitly built for long context. More tokens is not more memory. You cannot brute-force your way out of a coordination problem; you have to architect what each mind holds, forgets, and passes on.

Second, on the multi-agent side: a Berkeley group catalogued what goes wrong when you wire many agents together. Why Do Multi-Agent LLM Systems Fail? (Cemri et al., 2025) hand-annotated more than two hundred runs across seven frameworks and found that failures cluster into three buckets — specification (the roles and goals were under-defined), inter-agent misalignment (they talked past each other, dropped context, or stepped on one another), and verification (nobody checked the work). Read that list again. Not one of those is a model-capability problem. They are the exact problems a real company invents process to solve: job descriptions, standups, handoffs, code review. We are rediscovering organizational design, one painful agent trace at a time.

There is an old observation lurking here. Friedrich Hayek argued in The Use of Knowledge in Society (1945) that the fundamental economic problem is that knowledge never exists in one head — it is dispersed across many, each holding a fragment. Coordination is the work of putting those fragments to use without first gathering them into a single impossible mind. That is precisely the situation with a swarm of agents, each with its own slice of context. And Conway's law (1968) warns of the flip side: a system inevitably mirrors the communication structure of whatever builds it. Give your agents a bad org chart and they will build you a bad product. The structure is not a detail. It is the thing.

The company is the model

So here is the reframe I keep returning to. The interesting object is not the agent. It is the company the agents run inside.

If that is right, the moat is not the model — everyone rents the same weights — it is memory. A company that has operated for six months should beat one that started yesterday, for the same reason a six-year employee beats a new hire: it remembers what was tried, what failed, who owns what, and what the goal is. This is not hypothetical. The most striking demonstration I know is Generative Agents (Park et al., 2023), where twenty-five simple agents were given three things — a memory stream of everything they experienced, a reflection step to distill those experiences into higher-level beliefs, and planning that drew on both — and left to live in a small town. With nothing but a single seed ("one agent wants to throw a Valentine's party"), they spread invitations, formed relationships, and coordinated to show up at the right place and time. The intelligence that produced a functioning little society was not in any one agent's weights. It was in the memory architecture connecting them.

Turn that accumulated, structured memory into a substrate every coworker reads from and writes back to, and context stops resetting every session. It compounds. An organization that compounds context is, functionally, an organization that learns — without anyone retraining a model.

Where does the human stand?

Here is the part I care about most, because it is a human–computer interaction question before it is an infrastructure one: as the agents get more autonomous, where does the human go?

The honest answer is onto the loop, not out of it. You move from approving every step — human-in-the-loop — to supervising many agents at once and intervening on exceptions — human-on-the-loop. That shift is inevitable, because a person cannot babysit a swarm one click at a time. But it hides a trap that the safety literature named four decades ago. In Ironies of Automation (1983), Lisanne Bainbridge pointed out the cruel paradox of automating a process: the more the machine does, the more crucial — and the harder — the human's leftover job of monitoring it becomes. The operator is asked to stay vigilant over a system they no longer actively run, to catch the rare moment it goes wrong, with skills that atrophy precisely because the machine is usually right.

A dashboard full of green checkmarks is not oversight; it is the illusion of it. So the design problem — the one I think is worth a career — is to make autonomy legible enough that supervision stays real. Ben Shneiderman makes the case in Human-Centered AI (2020) that high automation and high human control are not opposite ends of one slider — you can, and should, design for both at once. That is the goal: not a config file you set and forget, but a control room — a seat where one person can watch an autonomous organization work, understand what it is about to do, and redirect it before it is too late. This is the real meaning of raising the bandwidth of human–AI collaboration: letting one human meaningfully steer many minds.

The one-person company is a coordination story

All of this lands on a prediction that sounds like hype until you trace the mechanism. Dario Amodei has put the odds of the first one-person billion-dollar company at something like 70–80% within a couple of years. The lazy reading is "AI replaced the employees." The accurate reading runs straight back through Coase.

Coase said the size of a firm is set at the margin where the cost of organizing one more task internally equals the cost of buying it on the market. Lower the internal coordination cost and that boundary moves — companies can do more, with fewer seams, before they fragment. Now imagine the coordination cost of adding a worker falling toward zero, because the worker is software that shares memory with every other worker instantly and never forgets the brief. Headcount decouples from output. The founder stops being a manager of tasks and becomes a designer of an organization — one whose employees happen to be software and whose org chart happens to be a program you can edit. The thing you build is no longer a product or even a team. It is the company itself, versioned and run.

That is the bet I am chasing: that the most important software of the next decade is not another agent, but the operating system the agents live inside — a place to belong, a memory that persists, coworkers to coordinate with, and a control room where a human can still steer. I did not get here from a grand thesis about AGI. I got here by building agents until it became obvious that the agent was never the hard part.

Attention was all you needed to make a model think. Coordination is all you need to make a company of them work. Let's see how far it can run.