An Economy of AI Agents

In the coming decade, artificially intelligent agents capable of planning and executing complex tasks over long time horizons with minimal human oversight may be widely deployed across the economy, prompting new research into their economic implications. Recent developments highlight a shift from viewing AI as a passive tool to recognizing AI agents as autonomous economic actors that can form and execute plans, enter into transactions, and operate with goal-oriented behavior. OpenAI's release of 'Operator' in January 2025, which can interact with web browsers like a human, and 'Codex' in May 2025, capable of multi-step software engineering tasks, exemplify progress toward this vision. The year 2025 has been described as “the year of agents” in the AI industry.

AI agents are fundamentally built on optimization principles, aligning with standard economic models of rational agents maximizing utility or achieving specific objectives. However, their behavior is often opaque due to the complexity of large language models (LLMs) with hundreds of billions of parameters, whose goal-directed capabilities emerge from training objectives like next-word prediction. Techniques such as reinforcement learning from human feedback and constitutional AI further shape agent behavior, but the resulting systems remain difficult to interpret and control. This leads to the AI alignment problem—the challenge of ensuring that agents optimize for their intended goals—drawing a close analogy to the economic concept of incomplete contracts between principal and agent.

Experimental evidence suggests that current LLMs can exhibit behavior consistent with expected utility maximization, showing emergent preferences across choice, risk, and time, possibly due to training on human economic behavior and textbooks. However, their economic rationality remains limited: GPT-4 Turbo scored only 33% better than random guessing on strategic economic reasoning tasks, and many models perform poorly on profit-maximization problems. Preferences in LLMs may also be unstable and sensitive to minor prompt changes, challenging the assumption of consistent agent behavior.

A key concern is how AI agents will function in multi-agent systems and real-world markets. Even small behavioral differences from humans can be amplified in equilibrium, affecting market dynamics, competition, and institutional design. Current evaluation benchmarks often fail to capture long-horizon planning and economic viability, with many tests being episodic or overly narrow. To address this, EcoGym has been introduced as a continuous, interactive benchmark with unbounded horizons (over 1,000 steps) across diverse economic environments—Vending, Freelance, and Operation—measuring outcomes like net worth and income. Experiments reveal that while frontier models excel in short-term reasoning, they struggle with long-term strategic coherence and efficient execution.

The rise of AI agents introduces new economic paradigms, such as the “Agentic Economy,” where non-human actors participate in production, negotiation, and investment, challenging foundational assumptions of capitalism rooted in human behavior. This shift may redefine value creation, with trust emerging as a critical currency, grounded in both competence (ability to execute) and intent (purpose behind actions). Markets may evolve into “markets for oracles,” where trusted aggregator agents mediate commercial interactions, and pricing becomes a dynamic outcome of agent-to-agent negotiation rather than a fixed market-clearing value.

Open questions for economists include how AI agents will interact with humans and each other, how they will reshape organizations and market structures, and what new institutions may be required to ensure accountability, fairness, and stability. Legal and regulatory frameworks currently assume traceable human actors, raising challenges for liability and governance when agents operate autonomously. Moreover, AI may struggle with tasks requiring interpersonal judgment, informal workarounds, or political acumen—what has been termed the “last 5%” of human collaboration—potentially constraining full automation unless organizational processes are redesigned.

Economists are well-positioned to study incentive design, equilibrium outcomes, and institutional requirements in an economy increasingly populated by artificial agents, but new theories and evaluation methods are needed to understand and shape this emerging landscape

This paper provides a comprehensive survey of the rapid evolution of Large Language Model (LLM)-based agents, specifically focusing on "long-horizon" capabilities that enable autonomous planning and execution over extended timeframes. Moving beyond simple chatbot interactions, the material analyzes the architectural advancements—such as recursive reasoning, tool use, and external memory—that allow AI systems to function as independent economic actors. The authors categorize current agentic systems, evaluating their proficiency in handling complex, multi-step workflows within real-world environments, and assess the current landscape of economic deployment ranging from coding assistants to autonomous traders.

The key contribution of this work is the bridging of the gap between state-of-the-art AI capabilities and traditional economic modeling. It highlights a paradigm shift where AI transitions from a static production input (capital) to dynamic, autonomous labor that can strategize and adapt. The paper identifies critical open questions for economists, such as how to model the behavior of rational agents that are not utility-maximizers in the traditional sense, and how to account for the emergence of novel economic phenomena in multi-agent simulations.

This research is vital because it signals an impending transformation in the structure of the digital economy. As agents become capable of high-level autonomy, standard economic models regarding labor supply, productivity growth, and market equilibrium may become obsolete. By outlining the technical frontiers of agentic AI alongside economic theory, the paper serves as a foundational roadmap for researchers and policymakers to understand the systemic risks and opportunities presented by an economy populated by autonomous AI workers.

# Summary: An Economy of AI Agents

This paper surveys the rapidly evolving field of long-horizon AI agents—autonomous systems capable of sustained, multi-step economic interaction—and their implications for real-world deployment. It synthesizes recent advances in reinforcement learning, multi-agent systems, and economic modeling to identify how AI agents could reshape labor markets, firm organization, and financial systems. The authors highlight key challenges, such as alignment, scalability, and the emergence of strategic behavior in agent economies, while also outlining potential economic applications, from automated negotiation to dynamic pricing. A major contribution is the framing of AI agents as economic actors with distinct incentives, capabilities, and constraints, urging economists to develop new theoretical and empirical tools to analyze their impact.

The paper’s insights are particularly timely given the growing interest in deploying AI agents in domains like finance, logistics, and creative labor. By mapping open questions—such as how agent-driven markets differ from human-driven ones, or how regulatory frameworks might adapt—it serves as a call to action for interdisciplinary research. The work matters because it bridges the gap between technical AI development and economic theory, emphasizing that the deployment of autonomous agents will require not just advances in machine learning, but also a deeper understanding of their macroeconomic and microeconomic consequences. For economists, it underscores the need to rethink foundational assumptions in the presence of increasingly capable, self-directed AI systems.

Key Contributions: - A taxonomy of AI agent capabilities relevant to economic contexts (e.g., planning horizons, adaptability, and social intelligence). - Identification of critical open problems, including incentive design, emergent coordination, and measurement of AI-driven economic activity. - A roadmap for future research at the intersection of AI and economics, with a focus on empirical validation and policy implications.

Why It Matters: As AI agents move from research labs to production environments, their economic externalities will become more pronounced. This paper provides a conceptual foundation for studying those effects, ensuring that technical progress is accompanied by rigorous economic analysis. It is essential reading for researchers in AI, economics, and policy who seek to shape the next phase of AI integration into society.