This isn’t a thought experiment. At the most recent Y Combinator demo days, startups are showing up with five times more revenue per employee than the cohorts from eighteen months prior. Companies that would have needed a head of marketing, a sales team, a customer support function, a finance person, and a handful of engineers are launching with two or three people and a budget line for inference costs. The work is still getting done. But it’s not the humans doing it anymore.
The default explanation for this is that AI makes people more productive. That’s true but incomplete. What’s actually happening is more structural: software is no longer a tool humans use to work faster. Software is becoming the work itself. The limiting factor for a company’s output used to be headcount. It’s becoming token usage.
To understand why this matters, why it isn’t just a cost story but a shape-of-the-entire-economy story, you have to understand why big companies existed in the first place.
Coase in 60 seconds
In 1937, an economist named Ronald Coase asked a question nobody had bothered with: if markets are the most efficient way to allocate resources, why do firms exist at all? Why don’t we all just freelance and buy everything on the open market?
His answer was transaction costs. Finding suppliers, negotiating contracts, writing agreements, enforcing promises, coordinating delivery, all of this friction made it cheaper to hire people and manage them internally than to buy every function on the open market. A firm grows until the cost of organizing one more task in-house equals the cost of outsourcing it. When internal coordination is cheaper than market friction, the firm expands. When it isn’t, the firm contracts.
That’s it. That’s the entire theory, and it has explained the shape of every company for almost a century. The reason your company has a legal department instead of hiring a lawyer per contract, a marketing team instead of briefing an agency per campaign, and a finance function instead of outsourcing every invoice is that Coase was right: the transaction costs of going to the market every time were too high.
The question is what happens when those costs go to zero.
What agents actually collapse
Coase identified several categories of friction that justify a firm’s existence. Agents are dismantling them one by one.
Search costs: Finding the right supplier, contractor, or service provider used to take days or weeks, calls, referrals, RFP processes, and comparison spreadsheets. An agent does this in minutes. It scans vendors, compares terms, checks reviews, evaluates fit against constraints, and presents a shortlist. The procurement process that used to justify an entire sourcing department becomes a five-minute conversation with an agent that already researched before you asked.
Negotiation costs: Every deal between two parties involves back-and-forth pricing, scope, timelines, and terms. This is why companies have account managers, sales teams, and partnership leads: humans negotiating with humans is slow and expensive. Agent-to-agent negotiation collapses this. When both sides of a deal are represented by software that can evaluate terms, propose counteroffers, and close agreements in seconds, the human cost of deal-making drops to near zero. The agent doesn’t need a meeting. It doesn’t need lunch.
Enforcement costs: Once a deal is made, someone has to make sure it’s honored, deliverables met, invoices paid, and quality maintained. This is why companies have project managers, vendor management teams, and quality assurance functions. An agent monitors continuously. It flags deviations the moment they happen, not at the next quarterly review. The enforcement layer that used to require a team becomes a background process.
Coordination costs: This is the big one, the category that explains why companies have middle management at all. In a traditional firm, information flows up from the people doing the work, gets synthesized and prioritized by managers, and flows back down as decisions. Every layer adds latency and distortion. The whole structure exists because humans can’t coordinate at scale without a relay system.
Agents don’t need a relay system. Information goes from source to decision-maker, whether it’s a human or an agent, without passing through five people who each add a day of delay and a layer of interpretation. The weekly status meeting, the alignment sync, the cross-functional review, these are all artifacts of a coordination bottleneck that agents simply don’t have. An agent can read every project update, every metric, and every customer complaint and synthesize them in seconds. The middle of the org chart was a routing layer. Agents don’t need routers.
What the firm becomes
When you collapse search, negotiation, enforcement, and coordination costs simultaneously, the firm doesn’t just get leaner. It changes shape entirely.
The 50-person company was never 50 people doing 50 different jobs. It was maybe 8–12 people doing the actual differentiated work, and 38–42 people searching, negotiating, coordinating, enforcing, relaying, reporting, and managing each other. The overhead wasn’t a waste. It was the cost of operating at scale when every connection between tasks required a human to bridge it.
Remove that cost and the firm contracts to its core: the people who do the things only humans can do. Judgment in novel situations. Relationships that depend on trust between specific people. Creative leaps that require taste, not optimization. Ethical decisions that carry real weight. Everything else, everything that’s repeatable, coordinatable, or monitorable gets handed to agents.
The resulting company looks something like this:
A core team of one to five humans. A fleet of agents handling operations, research, procurement, customer support, analytics, scheduling, compliance monitoring, content production, and financial management. The humans set direction, handle exceptions, and maintain the relationships that matter. The agents do the volume.
The org chart is flat, not because the founder read a blog post about flat hierarchies, but because the layers that used to sit between the founder and the work have no economic function anymore. You don’t need a VP of Marketing to manage a team of five marketers when an agent handles campaign execution, analytics, and optimization. You don’t need a Head of Operations to coordinate across departments when the agents already share context. The organizational structure simplifies because the friction that justified the complexity is gone.
A company’s “size” is no longer its headcount. It’s its agent capacity. A three-person company orchestrating fifty agents has the output of a fifty-person company with the overhead of a three-person company. The economics aren’t close. They’re an order of magnitude apart.
What breaks
If this reads like a frictionless future, it’s because I haven’t told you what goes wrong yet. Plenty goes wrong.
Liability has no clear owner: When a 50-person company’s employee makes a bad decision, there’s a chain of accountability: the employee, their manager, the department head, and the company. When a three-person company’s agent makes a bad decision, signs a contract with unfavorable terms, sends a misleading email to a client, or miscalculates a tax payment, who’s responsible? The founder who deployed the agent? The developer who built the model? The API provider whose output was wrong? The legal system assumes humans in the decision chain. The agentic firm has humans at the edges and software in the middle, and nobody has figured out where the liability sits.
Institutional trust doesn’t transfer to agents: Clients hire firms partly because the firm’s brand is a guarantee. When you hire McKinsey, you’re not hiring a specific consultant; you’re hiring the institutional quality control that McKinsey’s name represents. A solo operator with a fleet of agents doesn’t have that. The work might be just as good, but the trust infrastructure hasn’t been built yet. This is a real barrier, not a perception problem; it means the agentic firm will penetrate some markets (software, content, analytics) much faster than others (law, finance, healthcare) where institutional trust is load-bearing.
Quality control is unsolved at scale: A 50-person company has peer review, institutional knowledge, and senior people catching junior mistakes. The agentic firm has... what? This is the same eval problem I wrote about in “Code Nobody Read”. If agents are doing the work and there aren’t enough humans to review every output, how do you verify quality? The answer today is mostly “you don’t,” which works for low-stakes tasks and is terrifying for high-stakes ones.
The regulatory stack assumes employees: Employment law, corporate tax structures, insurance frameworks, governance requirements, all of it is designed for firms that employ humans. A company with three employees and fifty agents doesn’t fit cleanly into any of these categories. Are agents capital expenditure or operating expenditure? Does a company with one employee and the output of fifty need the same governance as a fifty-person company? Tax law in most jurisdictions hasn’t even begun to address this, and the answers will reshape the economics significantly.
Loneliness is a real cost: This one gets dismissed as soft, but it’s structural. Humans in firms don’t just coordinate; they motivate each other, correct each other, teach each other, and hold each other accountable through social pressure. A solo operator with fifty agents has none of this. The productivity is there. The resilience, the growth, the error correction that comes from human teammates that might not be. The agentic firm might be economically optimal and psychologically fragile, and nobody’s talking about this because it’s not a technical problem.
What Coase would say
Coase’s theory always had an implicit corollary that people overlook. He didn’t just explain why firms exist. He predicted that if transaction costs ever fell to zero, firms would dissolve into markets, meaning every task would be outsourced, every function contracted, every interaction handled through market mechanisms rather than internal hierarchy.
For ninety years, this was a thought experiment. Transaction costs never fell to zero. They fell slowly; the internet reduced search costs, SaaS reduced coordination costs, freelance platforms reduced hiring costs, and firms shrank slowly in response. But the firm never dissolved. The friction was always just low enough that having a team was still worth it.
Agents are different from every previous wave of cost reduction because they hit all four categories of transaction cost simultaneously. Not sequentially, the way previous technologies did. The internet reduced search but not coordination; SaaS reduced coordination, but not negotiation; and outsourcing platforms reduced hiring but not enforcement. Agents reduce all of them, at once, toward the same floor.
Coase’s thought experiment is becoming a business plan. The firm isn’t disappearing, but its default size is collapsing from “as many people as we can coordinate” to “as few people as we can’t replace”. Everything between those two numbers was overhead, and the overhead is being automated.
The solo-operator firm isn’t a lifestyle choice. It’s the logical conclusion of an economic theory that’s been waiting ninety years for the technology to catch up.