An Economy of AI Agents

Abstract In the coming decade, artificially intelligent agents with the ability to plan and execute complex tasks over long time horizons with little direct oversight from humans may be deployed across the economy. This chapter surveys recent developments and highlights open questions for economists around how AI agents might interact with humans and with each other, shape markets and organizations, and what institutions might be required for well-functioning markets.

Introduction. This chapter outlines the possibility of AI as economic agents and their attendant implications for markets, organizations, and institutions. We highlight what we see as overlooked questions we think economists are particularly well-positioned to answer. Our goal is to stimulate research in this area rather than to be comprehensive; the important implications of AI for labor markets and growth, for example, are well addressed elsewhere in the economics literature (see Comunale and Manera (2024) for a review).

1.1 A primer on AI agents. To date, economists have largely focused on AI as a tool to be incorporated into production, focusing on its adoption (McElheran et al., 2024), labor market impacts (Acemoglu et al., 2022; Noy and Zhang, 2023), harms from prediction (Acemoglu, 2021), and macroeconomic possibilities (Acemoglu, 2025; Korinek and Suh, 2024). But AI development has increasingly shifted to the goal of producing AI agents capable of taking in general instructions (“Go make $1 million on a retail web platform in a few months with just a $100,000 investment” (Suleyman, 2023)) and autonomously forming and executing complex plans that require entering into economic relationships and transactions. In January 2025, OpenAI released its first AI agent ‘Operator’ which can operate a web browser the same way humans do—by typing, clicking, and scrolling. In May 2025, OpenAI released ‘Codex’, an autonomous agent that can perform complex and multi-step software engineering tasks. 2025 was widely heralded in the AI industry as “the year of agents.” (Kim, 2025). As for expected technical trajectories, OpenAI (Metz, 2024) has partitioned AI development into five stages, with agents in Stage 3 (“AI systems that can spend several days taking actions on a user’s behalf”) and organizations in Stage 5 (“AI systems that can function as entire entities, possessing strategic thinking, operational efficiency, and adaptability to manage complex systems.”). How would an economy of AI agents function? To what extent do our models of humans predict the individual and collective behavior of artificial agents?

AI systems are fundamentally built on principles of optimization: the current paradigm builds agents to achieve objectives given available actions and information (e.g., to maximize the probability of winning a game of Go or completing a coding task). This coincides with the paradigm within economics and, on this view, AI agents are well-described by standard economic models.1 But it is also important to recognize the fundamental ways in which the methods of building AI systems can drive a wedge between the predictions of economic theory and the behavior of an artificial agent.

Although early AI systems were built on the basis of interpretable goals and algorithms, today’s AI agents are built using machine learning techniques that routinely render their goals and behavior opaque. How and why neural networks actually work is still largely mysterious: the large language models (LLMs) on which AI agents are currently built consist of hundreds of billions of parameters and their goal-oriented capabilities— solving math problems, identifying relevant answers to questions in a mass of text, scripting silly limericks—are a strange emergent property of a system trained merely to predict the next word in a sequence. Layered onto these base (“pre-trained”) models are a variety of finetuning steps, such as reinforcement learning from human feedback which trains the base model to choose outputs that earn higher predicted scores from human evaluators (Ouyang et al., 2022), and constitutional AI which trains the base model to choose outputs it predicts will better fit a set of principles supplied by the developer (Bai et al., 2022). Final delivery of outputs can be further modified by filters and (generally hidden) system prompts that are appended to user prompts. Other techniques for building or finetuning AI models include reinforcement learning in which neural networks are optimized to achieve designer-specified rewards. But even in such cases, model behavior remains—and even becomes more—unpredictable due to the complexity of reward specification. This problem has a close analogy familiar to economists: the unavoidable incompleteness of contracts (reward specification) between principal (designer) and agent (AI). As a result, even though AI agents are optimizers, we cannot be sure what they are optimizing. This is known as the AI “alignment problem” (Hadfield-Menell, 2021).

Recent experimental work finds that the current generation of large-language models (LLMs) exhibit behavior consistent with expected utility maximization (Chen et al., 2023; Mazeika et al., 2025; Kim et al., 2024; Fish et al., 2025; Raman et al., 2025). That is LLMs can exhibit emergent preferences and behave like textbook economic agents across domains of choice, risk, and time.

Related work. The foundations of neoclassical economics rest on theorems about the general equilibrium and welfare characteristics of markets populated by rational agents pursuing individual self-interest (Arrow and Debreu, 1954; Arrow, 1951). What happens to these predictions when market participants are not humans, but artificial agents optimizing in complex ways on goals supplied or developed during commercially-produced machine learning processes that are themselves subject to competitive dynamics? In this section, we highlight open questions around what the presence of AI agents in the economy implies for prices, equilibria, and welfare.

A natural benchmark is the pure exchange economy of Arrow and Debreu (1954). The celebrated welfare theorems guarantee competitive equilibria are Pareto efficient (first), and that all Pareto efficient equilibria are implementable via a suitable reallocation of initial endowments (second). AI agents are likely to substantially reduce transaction frictions as they begin to act as personal shoppers, performing market research, and checking prices autonomously—this pushes us toward the Arrow-Debreu world. Yet, the difficulties of perfectly specifying human preferences introduces a wedge between human preferences and AI choice. This wedge introduces two kinds of distortions. First, holding market prices fixed, the resultant bundle of goods purchased by the AI on behalf of humans might simply be suboptimal. Such distortions are as if human decision makers made mistakes. The second distortion is in general equilibrium: the mistakes introduced by AI consumption could decouple prices from preferences such that they no longer reflect relative wants. Such distortions are precluded if (i) markets are large so each consumer has small price impact; and (ii) AI mistakes are zero-mean and independent of each other. If either condition fails, however—e.g., AI agents could be systematically biased toward certain marketplaces or their choices might be manipulated by a third-party—then prices might fail their classic role of aggregating information (Mises, 1920; Hayek, 1945).2

Method. 2.2 Prices and market power. If AI agents begin to play a substantial role in consumption, this will have substantial implications for market power. The first possibility is that AI agents might influence prices by selectively recommending products to humans who make final purchase decisions. Ichihashi and Smolin (2023) develop a model where AIs might strategically bias its recommendations such as to drive down equilibrium monopoly prices. Second, AI agents might have lower search costs or be better informed about product offerings, both of which can intensify competition in product markets. Finally, if AI agents can make purchase decisions (and not simply recommendations), systematic distortions in AI choices could alter the shape of the demand curve. Dai and Koh (2024) analyze how such wedges between choices and preferences can generate positive pecuniary externalities via lower market prices. These distortions can—even net of AI mistakes—improve consumer welfare. More broadly, we think there is more work to do tracing out the distortion-price frontier to understand how equilibrium prices are shaped by distortions from AI agents’ consumption choices. While such wedges between human preferences and AI choice are inevitable (Hadfield-Menell and Hadfield, 2019), how we handle with them is a design choice. How do we want AI agents to fill in the gaps between underspecified preferences? When do want them to refrain and defer to humans? What are the ensuing equilibrium implications?

2.3 Search and matching. Beyond environments with prices, AI agents might also act on humans’ behalf in matching markets. An important question is whether and to what extent noise in AI agents’ representation of their human counterpart might generate inefficiencies. Liang (2025) develops a model of ‘AI clones’. In her model, AI clones have substantially reduced search costs—searching over an infinite population of other AI clones—but are imperfect representations of their human counterparts. She shows that as the number of dimensions grow large, even small representation errors can lead to (relatively) worse matches under the AI clones regime: humans are better off searching in person over just two other humans. Liang (2025) is primarily interested in upper-bounding the value of AIs as representations of humans. A related but distinct question is whether AIs as agents—searching autonomously on behalf of their human counterparts—might generate equilibrium congestion, and if this can erase gains from the lowered costs from AI search in low dimensions.3 2.4 Collusion and bargaining. Price setting on digital platforms is already driven by algorithms. A remarkable finding is that independent AI agents are able to collude on supracompetitive prices in repeated price-setting games (Calvano et al., 2020; Fish et al., 2024). There is also real-world evidence from the 2017 introduction of algorithmic pricing into Germany’s retail gasoline market (Assad et al., 2024).

AI agents might also bargain on behalf of humans, and potentially with each other.5 There are parallels with the classic insight of Schelling (1960) that delegating bargaining to agents with different incentives can deliver a strategic advantage. A common theme from the literature is that a principal (human) often wishes to delegate bargaining to an agent who is less desperate to reach agreement.6 When the agent is another human, there are typically practical constraints over the kinds of agent preferences the principal can induce. But such constraints are less severe with AI agents since their preferences can, in principle, be chosen quite flexibly (Conitzer, 2019). This might be formalized as a preference selection game in which in the first-stage humans choose the reward functions (preferences) of their AI agents, and in the second-stage these agents bargain over surplus.7 A dangerous possibility is that the flexibility to shape AI agents’ preferences can lead to a ‘race to the bottom’ and surplus destruction.

Discussion. across every stable equilibria—all economic activity is driven by AI producing AI and human consumption is driven down to zero.

The obstacles that prevent human firms from growing without bound seem intrinsic to humans but not to AI. For instance, human communication is inherently rate-limited so we ‘know more than we can tell’ (Polanyi, 1966). On the other hand, information can be transmitted and processed near-instantaneously between artificial agents. Further, (most) humans have an inherent dislike for work; not so with AI agents whose reward functions can ostensibly be designed to prevent shirking which renders monitoring and enforcement—either via fiat or contract—unnecessary. If AI agents can, in fact, coordinate and resolve incentive problems more efficiently than humans, this will have profound consequences for economy-wide industry structure. A useful taxonomy here is distinguish (i) economies of scale; (ii) economies of scope; and (iii) new industries which do not yet exist. We discuss each in turn.

3.4 Systemic fragility. Over the past decade, economists have analyzed how small shocks might be amplified and propagate across the economy. The increasing adoption of AI agents within the firm might exacerbate such fragility. A straightforward channel is that the errors introduced by AI agents might be more correlated than those of humans. This might arise because the same agent is ‘copied’ both within and between firms, inducing correlated mistakes that do not wash out in the aggregate. For instance, automated trading algorithms likely exacerbated the 2010 ‘Flash Crash’ that wiped out approximately $1 trillion over the span of 15 minutes (Kirilenko et al., 2017).17 Furthermore, how AI agents behave—especially in complex ‘out of sample’ environments—is still poorly understood. This poses a challenge for models of systemic fragility which typically start from a fully-specified model of how agents learn and optimize, then studies emergent behavior e.g., cascading financial or supply-chain failures (Elliott and Golub, 2022). How might we analyze an economy of opaque ‘black box’ agents in a ‘detail-free’ way? How should we robustly intervene to safeguard against fragility?

AI agents are built on foundation models (Bommasani et al., 2021), the most advanced of which are built inside private firms. This renders them doubly-inscrutable. As we have already emphasized, we do not understand why or how massive neural networks function, and the mapping from inputs (data and training procedures) to model outputs is mysterious and based largely on trial-and-error. And because frontier models are now trained with considerable secrecy within private labs, we don’t even know what goes into such models. Nor are smaller open-weight models good guides—they simply do not display the capabilities and behaviors of larger models (Kaplan et al., 2020).

This presents a serious challenge to our regulatory and legal institutions. At present, regulatory implications do not feature as a significant consideration in the legal design of the boundary of the firm. After all, regulators don’t need access to the internal processes of automobile or pharmaceutical manufacturers in order to assess their safety and performance—they can simply test the final products or draw on public domain science to evaluate them. But the massive AI models developed in commercial labs cannot be replicated and evaluated in government or academic labs: the costs of training are too high and evaluation requires access not only to model outputs but also inputs—their data and training procedures.

Conclusion. 5 Concluding remarks Silicon Valley promises us increasingly agentic AI systems that might one day supplant human decisions. If this vision materializes, it will reshape markets and organizations with profound consequences for the structure of economic life. But, as we have emphasized throughout this chapter, where we end up within this vast space of possibility is a design choice: we have the opportunity to develop mechanisms, infrastructure, and institutions to shape the kinds of AI agents that are built, and how they interact with each other and with humans. These are fundamentally economic questions—we hope economists will help answer them.