Coase’s Theory of the Firm Explains Multi-Agent AI
In 1937, Ronald Coase asked an embarrassingly simple question: if markets are so good at coordinating economic activity, why do firms exist at all?
If I need steel, accounting, software, advertising, and transportation, why not simply buy each service from the market whenever I need it? Why create a company, hire employees, establish managers, and replace market prices with instructions like “do this”?
Coase’s answer was transaction costs.
Using a market isn’t free. You have to find suppliers, communicate what you want, negotiate terms, verify the result, enforce agreements, and repeat much of the process the next time you need something.
Inside a firm, many of those transactions disappear. Instead of negotiating a new contract with a programmer every morning, I employ a programmer and tell him what needs to be done.
But firms aren’t free either. They have coordination costs: management, bureaucracy, meetings, reporting structures, information loss, bad incentives, and increasingly cumbersome decision-making as they grow.
So the boundary of a firm emerges from a tradeoff:
Use the market when the cost of transacting externally is lower than the cost of coordinating internally. Bring an activity inside the firm when the reverse is true.
Nearly ninety years later, this is a surprisingly powerful way to think about AI agents.
An AI agent has the same choice
Imagine an AI system needs research, programming, legal analysis, image generation, and financial modeling.
One architecture is essentially a firm. A central agent has a collection of subordinate agents, shared memory, common tools, established roles, permissions, procedures, and perhaps a hierarchy. The programmer agent doesn’t negotiate a contract every time it receives a programming task. The organization already knows that programming belongs to it.
But we could build the opposite architecture.
An agent needing some code could discover other agents offering programming services, request bids, inspect reputations, negotiate a price, purchase the work, evaluate the result, and move on. Next time it might hire a completely different agent.
That’s a market.
The interesting question is not whether one architecture is inherently better.
It is the same question Coase asked:
When is it cheaper for agents to coordinate internally, and when is it cheaper for them to transact externally?
AI changes both sides of Coase’s equation
This gets interesting because AI can radically reduce external transaction costs.
Software agents can search enormous markets nearly instantly. They can compare thousands of suppliers, negotiate machine-readable contracts, make micropayments, inspect reputations, run automated evaluations, and switch suppliers with little emotional or institutional friction.
Activities that required employees because contracting was cumbersome may become cheap to purchase dynamically.
That pushes toward smaller firms and more markets.
But AI simultaneously attacks the other side of Coase’s equation.
Agents can also reduce internal coordination costs.
A human manager can only supervise so many people. Organizations therefore accumulate layers:
CEO VP Director Manager Worker
Each layer exists partly because human attention and communication bandwidth are scarce.
An AI coordinator can potentially manage hundreds or thousands of specialized agents, distribute tasks, maintain state, monitor outputs, run evaluations, and propagate information automatically.
That pushes in exactly the opposite direction:
larger organizations and more internal coordination.
So the simplistic prediction that “AI will make companies smaller” doesn’t necessarily follow.
AI lowers the cost of markets.
AI also lowers the cost of firms.
What matters is which cost falls faster for a particular activity.
The firm becomes software
There’s another difference between human and AI organizations that Coase couldn’t have anticipated.
Human organizations are difficult to reconfigure.
Hiring takes time. Firing is costly. Employees accumulate relationships and institutional knowledge. Departments develop political interests. Reporting structures become entrenched.
An organization of agents can be far more fluid.
A coordinating agent might instantiate twenty specialists for five minutes, dissolve the organization when the task is finished, and create a completely different hierarchy for the next problem.
The boundary between market and firm therefore stops looking like a binary distinction.
We can vary:
- authority
- incentives
- shared knowledge
- shared capital
- identity
- reputation
- duration
- decision rights
- risk
- access to organizational memory
An agent might be deeply integrated into an organization for thirty seconds.
Another might remain independent for years while repeatedly selling services to the same organization.
We get something closer to a continuous space of organizational forms than a clean choice between employee and contractor.
AI adds another particularly strange dimension: cognitive externalization.
A human employee carries much of his knowledge inside his head. An AI organization can move enormous portions of cognition outside any individual agent and into shared prompts, memory, workflows, knowledge graphs, evaluations, permissions, tools, and decision procedures.
The intelligence of the organization increasingly resides in the organization itself.
The organizational operating system may become the asset
This changes what it means to own a productive organization.
Suppose every company has access to approximately the same frontier models.
Company A and Company B can both hire the same artificial programmer, researcher, lawyer, and analyst.
Where does competitive advantage come from?
Increasingly, it may come from the architecture surrounding those models:
What does the organization know?
Which agent gets which problem?
What information does each agent receive?
Who is allowed to make which decisions?
How are results evaluated?
Which failures cause escalation?
What gets remembered?
What gets forgotten?
When does the organization use an internal agent, and when does it go to the market?
The valuable artifact begins to look like an organizational operating system: workflows, knowledge, evaluations, standards, permissions, customer context, accumulated decisions, institutional memory, and the rules by which intelligence is allocated.
The models themselves may increasingly resemble interchangeable labor.
The organization that coordinates them may not be.
The boundary becomes programmable
This suggests a useful way of thinking about future AI-native companies.
For every task, an organization could continuously ask:
Should I do this internally?
Should I instantiate a temporary specialist?
Should I delegate it to a persistent agent that knows my organization?
Should I purchase the result from an external agent?
Should several independent agents compete and let an evaluator choose the winner?
The answer doesn’t have to be encoded permanently in an org chart.
It can be computed.
The boundary of the firm becomes programmable.
And once that happens, Coase’s theory stops being merely descriptive economics.
It becomes an engineering principle.
A multi-agent architect is effectively deciding which transactions should occur inside an organization and which should cross its boundary. Shared context, API calls, inference costs, verification, latency, reputation, communication overhead, memory, and delegation failures all become forms of transaction or coordination cost.
The architecture of an AI system is therefore partly an economic problem.
From economics to engineering
This may be one of the stranger consequences of AI.
Economists developed theories of firms, markets, incentives, specialization, information, and prices because humans needed ways to coordinate millions of independent minds.
Now we are beginning to manufacture minds.
And as soon as we connect those minds together, many of the same problems reappear.
Should they specialize?
Should they trade?
Should they share memory?
Should they form hierarchies?
Who should have authority?
How should scarce compute be allocated?
When should an agent delegate?
When should an organization use a market?
The remarkable thing about Coase’s theory is that we may no longer have to merely observe the answer in human institutions.
We can build the institutions ourselves.
The theory of the firm is becoming a theory of multi-agent architecture.