Theory of the Firm
Theory of the firm, in the Coasean sense, answers why firms exist at all in an economy supposedly coordinated by prices. Ronald Coase’s “The Nature of the Firm” (Economica, November 1937) opens with the puzzle: inside a firm, resource allocation is not carried by the price mechanism — “If a workman moves from department Y to department X, he does not go because of a change in relative prices, but because he is ordered to do so” — so why do these “islands of conscious power” (Coase quoting D. H. Robertson, “like lumps of butter coagulating in a pail of buttermilk”) form in the ocean of market exchange?1
The answer: the cost of using the price mechanism
Market exchange itself costs something, and that is Coase’s answer. Discovering the relevant prices, negotiating and concluding a separate contract for each transaction, and committing to terms that cannot anticipate the future all push activity out of spot markets. A firm replaces a series of contracts with one: a factor of production “agrees to obey the directions of an entrepreneur within certain limits,” with the details decided later inside those limits. Long-term contracts for services generate firms more readily than commodity purchases, because what is being bought can be stated only in general terms and the buyer wants the right to direct the details.
Boundaries
Firms expand until organizing one extra transaction inside costs what the same transaction costs on the open market or in another firm. Diminishing returns to the entrepreneur function, misallocation as firms grow, and the rising supply price of small-firm advantages supply the counterpressure; on the margin businessmen “will be constantly experimenting, controlling more or less.” Regulation tilts the margin: a sales tax on market transactions, quotas, and price controls all subsidize organizing inside the firm and tend to enlarge firms independently of efficiency. Coase’s reply to Knight’s uncertainty theory separates payment from direction — much work is done to contract for a guaranteed result with no direction involved — and makes direction the essence of the firm’s legal analog, the master-and-servant relationship.2
Application: unbundling and agent markets
Krishnan’s 2026 experiment takes Coase at his word about falling transaction costs: as they fall, firms unbundle, and the resulting mini-firms have to coordinate “by planning, or by markets.”3 He runs the choice on LLMs: 15 hand-written tasks (five coding, five reasoning, five synthesis) under three topologies — a strong model alone; a hub that decomposes, delegates, red-teams, and revises; and a market where three models bid per task, with a winner picked and reputation updated across the run.
The market averaged 7.2 out of 10 at a total cost of $1.34, solo averaged 7.2 at $1.69, and hub-spoke averaged 6.7 at $5.33. Task shape decided the winner: coding tasks holding global state (one data structure for the cache, one race list for the bug hunt) favored the solo model; a cleanly decomposable refactor favored the hub; brittle reasoning problems, where independent attempts plus failure detection and retries pay, favored the market. Krishnan’s summary: “Markets beat managers when the value of independent retry exceeds the value of orchestrated coherence.” The sample is small (n = 15 tasks) and the author labels the lesson accordingly.
MarketBench provides the measurement underneath: agent bidders are badly calibrated on their own success probability and token usage — “some models are overconfident (Gemini), some are underconfident (GPT)” — so auctions diverge from full-information allocation. Agents remain useful together anyway when the diversity premium (getting another model to try after a failure) is large enough.4
Open questions
- Whether model self-knowledge improves with training or elicitation (MarketBench’s intervention moved bids by giving agents information), or whether harness-level routing keeps the hub role dominant in practice despite its measured cost.
- Which transaction costs actually fall as agent tooling improves: decomposition and recomposition overhead (the hub’s hidden tax) or bidding and retry overhead (the market’s).
- Whether current multi-agent harness architectures — “all versions of hub-spoke models” in Krishnan’s account — are an equilibrium or a default.
Cross-domain connections
- knowledge-problem — the complementary question: why prices coordinate dispersed knowledge well when they do
- interstate-energy-permitting — federal condemnation replaces repeated state-by-state acquisition, while route approval and court-set compensation remain outside any firm-like hierarchy
- bullshit-jobs — the managerial layer that falling coordination costs should thin, and Graeber’s evidence that it has not
- invisible-hand — the market side of the coordination choice, in Smith’s actual formulations
Sources
- R. H. Coase 1937 — The Nature of the Firm
- Rohit Krishnan 2026 — Why Coase needs Hayek
- Andrey Fradkin and Rohit Krishnan 2026 — MarketBench: Evaluating AI Agents as Market Participants