What Coase and Simon would say about an agent

Why do firms exist? Ronald Coase asked the question in 1937 and answered that organising work inside a firm is sometimes cheaper than buying it on a market, because markets have costs: finding, negotiating, enforcing. Herbert Simon, in 1951, added the mechanism. An employment contract is not a purchase of specified work. It is an agreement by the employee to accept the employer's authority within a zone, in exchange for a wage, with the details to be filled in later. The firm is a bundle of such open-ended agreements.

Both accounts assume the thing being organised is a person. Coase's transaction costs are the costs of dealing with other people's self-interest. Simon's zone of acceptance is the range of orders a person will obey without renegotiating. Oliver Williamson later built an entire theory of organisation on two human traits: bounded rationality and opportunism. The firm, in this tradition, is a device for getting cooperation out of people who are limited and partly selfish.

Now introduce an agent. It has no self-interest in Williamson's sense; it does not shirk for its own benefit. It has, instead, a different and in some ways worse property: it does what it was trained to do, which is not always what it was told to do, and the gap is opaque. It has no zone of acceptance in Simon's sense, because it accepts everything and understands only some of it. Its transaction costs are near zero for the second copy and unbounded for the first mistake. The classical theory does not say the agent is capital or labor. It says nothing, because it was not built for an input that is neither limited in the human way nor obedient in the machine way.

Hadfield and Koh, in a 2025 chapter prepared for a handbook on the economics of transformative AI, make essentially this point: economic models of firms and markets cannot simply be relabelled to fit AI agents, because opaque training, alignment gaps and multi-agent dynamics introduce problems the models never contemplated. Ide and Talamàs, in a 2025 paper in the Journal of Political Economy, take the first formal step, placing autonomous and non-autonomous AI inside a model of knowledge hierarchies and showing that autonomy changes who becomes a manager and who becomes a worker. That is the frontier, and it is new.

What does this mean for a practitioner who cannot wait for the theory? It means recognising that the organisation now contains a kind of member its own rules were not written for, and that the rules will be rewritten by default, through a thousand local decisions, unless someone rewrites them on purpose. Simon's insight is the useful one here. The employment relation worked because authority was bounded and the bounds were understood by both sides. The agent needs an equivalent: an explicit boundary of what it may do alone, what with review and what never, owned by a person who understands the work. That is the deployment record I have described elsewhere in this series, and it is Simon's zone of acceptance, rebuilt for an entity that cannot negotiate it.

Coase's insight is useful too. The firm exists where internal organisation beats the market. For agents, the choice is already being made: execution is increasingly bought from vendors, while verification, the part that cannot be bought because it requires knowing the work, stays inside. That is a new firm boundary forming, and nobody is drawing it deliberately. The Chief Workforce Officer is, in Coase's terms, the person who decides where the boundary of the firm's productive capacity runs, and in Simon's terms, the person who writes the zone of acceptance for the members who cannot write their own.

Neither economist imagined the role. Both supplied the reasons it is needed.