Agents cheapen coordination in two directions. That does not dissolve the firm. It forces a split.
A company exists, in the old account, because it is cheaper to coordinate work inside a boundary than across a market. That was Coase’s observation in The Nature of the Firm. It was a statement about cost, not about identity.
When the cost of coordination falls far enough, the reason for the boundary changes; the legal shell can remain. The org chart need not.
The usual next sentence is the 70% question: is there a high-margin line two people and an agent stack could replicate in sixty to ninety days? That question is useful as a threat. It is the wrong question for design.
Agents do not lower one cost. They lower two, and the two point in opposite directions.
The split
Albert Azout states the mechanics cleanly:
Inside the firm, with regards to internal transaction costs, they automate search and summarization, structure handoffs, and perform continuous checks against rules and metrics, thereby lowering and flattening the marginal cost of organizing work and mitigating bureaucracy. Outside the firm, agents reduce external transaction costs by making it easier to discover, contract with, and monitor external partners, which pushes in the opposite direction toward more modular, plug-and-play relationships rather than monolithic, vertically integrated organizations.
That is the split. Cheaper internal coordination can make the firm larger — more workflows can sit under one roof before bureaucracy eats the gain. Cheaper external coordination can make the firm thinner — search, bargaining, contracting, and monitoring, the tax Coase said you pay when you go outside, can be performed at machine tempo.
The NBER / “Coasean singularity” line of work puts the market side this way:
The fundamental economic promise of AI agents lies in their ability to dramatically reduce transaction costs — the expenses associated with using markets to coordinate economic activity.
And:
The activities that comprise transaction costs — learning prices, negotiating terms, writing contracts, and monitoring compliance — are precisely the types of tasks that AI agents can potentially perform at very low marginal cost.
If that were the whole story, the prediction would be disintegration: much of what firms do would no longer need to happen inside them. Herbie Bradley reads the same theory and draws the opposite arrow. Continual learning and local process knowledge, he argues, reward the firm that already has the data:
I believe that the future of AI agents are just another step towards larger firms and greater market concentration, and (as many investors have observed), AI makes the economy’s power-law sharper.
Both readings can be true at once. They apply to different layers of the same company. Execution can modularize. The shell can thicken.
Gillian Hadfield and Andrew Koh ask the question the slide decks skip:
What happens to the boundary of the firm if significant numbers of transactions are carried out by AI systems?
Their working observation is that many limits on firm size were human limits — rate-limited speech, tacit knowledge that will not travel, dislike of monitoring. Those limits are not intrinsic to agents.
That does not mean the legal person evaporates. It means the economic rationale for the boundary and the legal rationale can come apart, and someone has to put them back together on purpose.
The firm is not a production function.
Nicolai Foss is useful here because he refuses the slogan. Coase, he reminds us, never said firms exist because talking is expensive.
Ronald Coase never argued that firms exist primarily because communication is expensive. Rather, he argued that firms exist because using markets is itself costly.
Williamson’s follow-on is the right instrument:
given a particular transaction, which governance structure minimizes the total costs of carrying it out?
Firms and markets are alternative governance structures, not alternative factories. Hierarchy still wins where adaptation is continuous, investments are specific, and contracts cannot name the future. Foss’s list of what remains hard to write down is almost a specification of the Evolut stack:
Decisions concerning technological direction, organizational design, capability development, resource orchestration, and the integration of proprietary knowledge are difficult to specify contractually because the relevant objectives themselves evolve as learning takes place.
So the interesting move is not “kill the firm” or “keep the org chart.” It is discriminating alignment: which transactions now belong in a market of agents, and which still have to sit inside a named legal person who can be sued, insured, and held to a purpose.
Two boundaries, not one
A later paper on agentic ecosystems draws a distinction that should be on every architecture diagram.
The interface boundary is the location of user interaction, task initiation, and workflow routing. The accountability boundary is the location of responsibility, evidence, review, signoff, and post hoc defensibility.
Agentic orchestration can move the first boundary without moving the second.
That is the error most “AI-native” programs make. They make a capability callable and think they have moved the firm. They have only moved the front door.
Making a capability callable is not equivalent to transferring the accountability boundary. Callable interfaces can expand markets when they allow agents to initiate, route, and retrieve work. They become strategically dangerous when they allow evidence, review, signoff, lineage, or responsibility records to be created or controlled outside the firm’s governed system.
Coasean logic correctly predicts cheaper assembly. It does not predict that liability, lineage, and signoff will follow the API.
AI-enabled capabilities whose outputs require evidence, review, signoff, or assignable responsibility may retain integrated accountability boundaries even when their technical interfaces become modular.
The twin built at the edge is not interesting because it is faster. It is interesting because it lets you choose which boundary moves.
What still has to sit inside the legal person
If coordination inside the firm becomes nearly free, the firm is no longer justified by its org chart. It is justified by four things that do not automate cleanly, and that a market of agents cannot hold in your name.
Liability. Someone has to be the residual claimant when an agent files, leases, launches, or pays. Hadfield and Koh note that new liability and agency rules for agents will be necessary; until those rules exist, the default is the company that deployed the stack. An interface you do not control is an action you may still own.
The data loop. Proprietary intelligence — your filings, your exceptions, your overrides, the cases that taught the model what “good enough” means here — is the only compounding asset that is yours. Bradley’s concentration argument is an argument about this loop, not about headcount. If the learning happens in a vendor’s swarm, you have rented a factory and sold the scrap.
Purpose as protocol. A poster on the wall is not a constraint. A protocol is. Agents will optimize the score they can see. The legal person is the place where purpose is written tightly enough to bind a stack: what we will not sell, whom we will not steer, which experiment is forbidden even if the cycle time looks good.
Judgment that cannot be reduced to a score. As knowledge becomes cheap, judgment becomes the scarce resource. Hierarchy does not disappear; it concentrates responsibility for decisions that cannot be written as rules. That is humans above the loop. Yes or no, and the reasons.
What does not
What need not survive inside the shell is coordination-as-management: the routing, drafting, reconciling, scheduling, and checking that made a firm feel like a firm. Those jobs can become a market of agents — yours, a supplier’s, a customer’s — so long as the four things above stay named.
Azout’s warning belongs here. Agents do not delete coordination costs. They redistribute them:
AI does not eliminate coordination costs—it redistributes them.
Integration debt, agent sprawl, opaque chains, audit trails that cannot be reconstructed — that is the new middle layer. If you outsource execution and also outsource observability, you have not modularized. You have gone blind.
What should become a market of agents?
A workflow belongs in a market of agents when three conditions hold at once.
- Search is cheap. Counterparties, prices, and capabilities can be discovered without a standing relationship.
- The contract can be written. Inputs, outputs, service levels, and failure modes can be named in advance — or renegotiated by agents faster than the exception matters.
- Enforcement is cheap. Compliance can be monitored, harm can be priced, and rollback does not require a committee that remembers why the work started.
The Coasean singularity paper is explicit about the expansion of what markets can even be:
By lowering the costs of preference elicitation, contract enforcement, and identity verification, agents expand the feasible set of market designs.
That is the right use of agents across the boundary: not “replace the vendor manager with a chatbot,” but make discovery, terms, and monitoring cheap enough that you no longer need to employ a department to perform the market.
The test is not whether an agent can do the task. The test is whether the governance structure of a market now fits the transaction better than the governance structure of a hierarchy.
Foss’s formulation is the adult version:
The interesting question is no longer whether AI will replace firms with markets. The more interesting question is which kinds of transactions become better candidates for markets, which remain more efficiently organized within firms, and which migrate toward increasingly sophisticated hybrid forms.
The twin is a hybrid on purpose. It runs inside the firewall. It reports to the chief executive. It uses a market of agents for the work that no longer justifies a desk, and it keeps the accountability boundary where the legal person can still answer.
A simple test
For each high-throughput workflow, ask only this:
Would we still own this workflow if search, contract, and enforcement were cheap?
If the honest answer is no — if we own it today because finding a counterpart is slow, because the contract would take six months, because no one outside can be trusted to notice drift — then the workflow is a candidate for a market of agents. Rebuild it at the edge. Measure cycle time, quality, cost, and risk against the old path. Deprecate what loses.
If the honest answer is yes — because the output must be defensible, because the learning is ours, because the purpose forbids the locally optimal act, because someone has to be sued in a name that exists — then the workflow stays inside the legal person. Agents may still execute. They do not get to hold the evidence, the loop, or the veto.
Interface can move. Accountability should not wander.
evolut_ing, as we use the word, is the act of taking ownership of one’s own consciousness to see a larger picture. Applied to a firm, the picture is not “the company is dead” and not “the org chart is sacred.”
The picture is the split: keep the judgment, keep the purpose, keep the learning that is yours, and stop asking a structure built for yesterday’s coordination tax to administer work that a market of agents can now clear.
Evolution before a revolution occurs. Draw the two boundaries on purpose.