The Coasean Singularity? Demand, Supply, and Market Design with AI Agents — Shahidi, Rusak, Manning, Fradkin, Horton (2025)

"The Coasean Singularity? Demand, Supply, and Market Design with AI Agents" is a 2025 chapter for the NBER Economics of Transformative AI volume by Peyman Shahidi (MIT), Gili Rusak (Harvard), Benjamin S. Manning (MIT), Andrey Fradkin (BU & MIT IDE; currently employed by Amazon), and John J. Horton (MIT & NBER). It is a companion to Hadfield & Koh's "An Economy of AI Agents." The chapter formalizes the economics of AI agents as market participants, framing the transition through the transaction-cost lens of Coase (1937) and developing implications for demand, supply, market equilibrium, market design, and regulation.

Core thesis

The chapter argues that AI agents will reduce the transaction costs that Coase (1937) identified as the central force shaping firms and markets — the costs of learning prices, negotiating terms, writing contracts, and monitoring compliance. Once agents can perform these tasks at low marginal cost, the authors hold that the make-or-buy boundary shifts and new "agent-first" market designs become feasible.

Demand for AI agents (Section 2)

Demand for agents is characterized as derived demand: humans do not enjoy watching an agent compile gas-grill price lists but delegate to achieve a market outcome at lower effort. The chapter identifies two adoption channels:

1. Substitution for human intermediaries (realtors, headhunters, brokers, financial advisors) — agents convert costly intermediation tasks (search, screening, quoting, negotiation, scheduling) into compute and API calls.
2. Enabling previously infeasible undertakings — by lowering exploration cost, agents expand the set of options worth pursuing (for example, diagnostic triage for home repairs, bespoke software scripts).

As an adoption heuristic, the authors expect agents to land first in markets where human agency is already common: high-stakes settings with a vast counterparty space, high evaluation effort, information asymmetries, and experience asymmetries. Examples given are real estate, job search, used cars, dating, and freelance hiring (LinkedIn, Upwork, Tinder, Zillow, Airbnb).

Designing AI agents (Section 3)

The core design problem is the alignment problem: eliciting principal preferences and ensuring the agent honors them. The chapter distinguishes two failure modes:

• Articulation limits — the principal cannot fully or consistently specify their preferences (for example, a full ranking of the Netflix catalog).
• Synthesis limits — the agent misinterprets what is stated (hallucination, inaccurate inference).

Preference dimensionality is treated as a key variable. Selling a house has low-dimensional preferences (price × time-to-sale), while buying has high-dimensional preferences (location, schools, safety, layout, age, resale). Higher-dimensional preferences amplify alignment errors and slow agent adoption (Liang 2025).

The authors also introduce meta-rationality: agents must learn not only how to act but when to act autonomously versus defer to principals, and must be both rational and robust to adversarial manipulation (analogous to "black-hat" SEO for LLM-consumed content).

Supply of AI agents (Section 4)

Production economics

The chapter identifies two provider types: frontier model builders (Anthropic, OpenAI) and agent providers using others' foundation models (Decagon, Harvey, Sierra). Training carries high fixed costs against the mostly variable cost of operation. The authors expect concentration in agent providers to mirror concentration in foundation models if vertical integration proves essential.

Pricing

Software is copyable at zero marginal cost, which the authors expect to diminish pricing power relative to human agents (who take a percentage of a transaction). A counter-force is that performance scales with compute, so prices may scale with compute allocated, possibly proportional to transaction stakes. The chapter lists likely pricing outcomes: free plus advertising (the search-and-social model), bundling (with phones, delivery), tiered freemium, and two-part tariffs (subscription plus per-prompt or per-token charges).

Consumer-facing typology — the four agent archetypes

Two orthogonal axes — ownership (Bring-Your-Own vs. Bowling-Shoe) and specialization (Horizontal vs. Vertical) — give a 2×2 archetype table:

The authors frame a trade-off between the two ownership models. Bring-Your-Own favors user alignment and portability but pays in maintenance, throttled access, and weaker first-party data. Bowling-shoe favors low friction and rich integrations but creates self-preferencing and lock-in risk. From the platform's view, Bring-Your-Own reduces costs and liability but sacrifices user-experience steering, while bowling-shoe maintains control and monetization paths but requires hosting and compute investment and may suppress consumer demand because of alignment concerns.

Equilibrium effects under the status quo (Section 5)

The chapter analyzes several equilibrium effects of capable agents:

• Behavioral-bias erosion. Markets where firms exploit bounded rationality (for example, suboptimal phone-contract selection per Grubb 2009) lose the rent extraction, as rational agents browsing on behalf of users resist nudges and puffery.
• Search-cost reduction, which improves allocative efficiency.
• Bargaining transformation. Classic bargaining models assume impatience because human time is scarce; for agents the binding constraint is compute. Agents can negotiate earlier (for example, starting Summer-2027 rentals in January 2026), hold many concurrent negotiations, and persist longer.
• Personalized pricing. Agents reveal preferences accurately to trusted intermediaries, giving firms richer data and enabling personalized pricing strategies that may improve discovery but redistribute rents toward firms unless agents strategically withhold.

The authors also identify counter-trends. Firms may respond to capable agents with sophisticated price obfuscation tactics (Ellison & Ellison 2009). When products are differentiated, lower search costs can lead to higher prices and greater dispersion (Ellison & Ellison 2018, Kaye 2024), which agents may exacerbate by enabling finer-grained matching. And congestion externalities arise as AI-generated cover letters flood employers (Kessler 2025; Wiles & Horton 2025), raising screening costs and degrading match quality.

Market design for AI agents (Section 6)

Identity (6.1)

The chapter expects most internet activity to originate from agents that can mimic humans, making Sybil attacks harder to detect. It describes two solution classes: walled-garden designs with login gates (imperfect, since bots launch after login and banned users create new accounts), and proof-of-personhood systems (for example, the World Foundation iris biometric), which are privacy-preserving but require restructuring and adoption. The authors expect identity to combine with verified credentials (digital government IDs revealing select demographics) and reputation mechanisms, enabling new market interactions such as targeted deals and small-purchase negotiations.

Platform-design changes (6.2)

The chapter anticipates three platform-design shifts. Agents acting as content filters review incoming streams and selectively transmit high-utility content, which challenges advertising- and engagement-driven platforms. Agents acting as content creators drive the marginal cost of posting, messaging, and applying toward near-zero, creating flooding incentives and pushing platforms to restore costly signaling (for example, per-post fees) or to verify human authorship. On infrastructure, a surge in agent-driven HTTP traffic shifts costs to website owners; Cloudflare's "pay-per-crawl" (Allen & Newton 2025) is described as the Pigouvian response, letting website owners charge agents for crawling. The authors expect new conventions akin to robots.txt to define agent permissions, with agent-first interfaces carrying machine-readable pricing and consent signals emerging alongside human-oriented pages. Amazon v. Perplexity is cited as a litigation manifestation of this fight.

Previously impractical mechanisms now feasible (6.3)

The chapter argues that agents make several previously impractical market mechanisms feasible:

• Deferred-acceptance algorithm (Gale & Shapley 1962) — guarantees stability and strategy-proofness but requires comprehensive preference rankings, historically infeasible to elicit at scale. Agents that can parse natural-language preference paragraphs (Rusak et al. 2025) make deferred acceptance practical for labor markets, dating, and similar settings.
• Privacy-preserving inquiry mechanisms — agents can pose sensitive questions (for example, about maternity leave policies or salary expectations) under precommitted protocols, separating sensitive inquiry from signaling concerns.
• Strategic anonymity — illustrated by Disney's anonymous land purchases for Walt Disney World, which the authors expect to become routine for agent-mediated transactions, leaving open whether multi-persona linkage to single individuals should be allowed (Buterin 2025).

Regulation (Section 7)

The chapter surveys four regulatory issue areas:

1. Market power — concentration among compute-, data-, and distribution-controlling firms, with antitrust scrutiny of frontier labs (DOJ Google decision, April 2025). The authors hold that regulation must preserve interoperability and prevent exclusionary practices while guarding against capture.
2. Autonomy and liability — the choice between negligence and strict liability for principal versus developer. The EU Product Liability Directive (2024/2853) explicitly extends liability to digital goods, software, and ongoing updates, adapting traditional tort frameworks rather than building new ones. (See AI Liability.)
3. Security and privacy — jailbreaking persists, and agents may pull data across contexts, retain it, or infer it from benign traces; CCPA and GDPR will need adaptation for generative AI and automated decision-making.
4. Data rights and platform access — training-data scraping (NYT v. Microsoft) extends to deployment-time scraping by external agents, with the contested question being whether platforms must license access or be required to interoperate.

Concepts introduced

The chapter is an economic-theory-grounded treatment of the agentic economy and introduces or develops several terms:

• The Coasean transaction-cost framing of the agentic transition (Coase 1937 applied to AI agents).
• The Bring-Your-Own versus Bowling-Shoe distinction as the structural debate over agent ownership and portability (illustrated by Cursor refusing Microsoft's Bring-Your-Own plus bowling-shoe acquisition, and Anthropic's MCP as a Bring-Your-Own-enabling protocol).
• Pay-per-crawl as the Pigouvian market response to crawler externalities.
• Specification hazard, drawn from the Imas/Lee/Misra 2025 companion paper, as the principal-agent informational problem.
• Agent-first markets (deferred acceptance, privacy-preserving inquiry, agent-only storefronts) as a design frontier.

Relationships

• supports: Agentic AI — economic-theory grounding for the agentic transition
• supports: Principal-Agent Problem Applied to AI — formalizes the alignment + specification problem
• supports: Agentic Interactions — Imas, Lee, Misra (2025) — explicitly cited as the demand-side companion (Shahidi et al. 2025 references)
• supports: AI Agentic Browsers — provides the economic model for pay-per-crawl, agent-first interfaces
• supports: Agent Supply Archetypes (BYO/Bowling-Shoe × Horizontal/Vertical) (new) — original BYO/Bowling-Shoe × Horizontal/Vertical taxonomy
• supports: Pay-per-Crawl (Pigouvian Pricing of Agent Traffic) (new) — economic justification (Pigouvian framing)
• supports: Transaction Costs (Coase) and the Agentic Economy (new) — Coasean lens on the agentic economy
• related: Amazon v. Perplexity AI — concrete instance of the agent-platform access fight
• related: NYT v. Microsoft, OpenAI et al. — cited as the training-data analog of deployment-time platform access
• related: Algorithmic Pricing and Antitrust — Calvano et al. 2020 on agent pricing collusion is referenced
• related: John J. Horton (new — author and lead empirical researcher in agentic markets)