An Economy of AI Agents (arxiv.org)

Andrew Koh’s new arXiv chapter, “An Economy of AI Agents,” surveys how autonomous, long‑horizon AI agents—systems able to plan and execute complex tasks with minimal human oversight—could be deployed across industries over the next decade. The paper synthesizes recent technical advances (e.g., multi‑step planning, reinforcement learning and multi‑agent coordination) and translates them into economic questions: how agents will interact with humans and each other, reshape markets and organizations, and what legal and institutional scaffolding will be needed to keep markets well‑functioning. The significance for the AI/ML community is twofold: first, agent capabilities will create new incentive, information and strategic interaction problems (principal–agent gaps, coordination failures, externalities, market power and systemic risk) that require technical and mechanism‑design responses; second, studying these phenomena demands new benchmarks, multi‑agent simulations, empirical measurement tools, and interdisciplinary work with economists and policymakers. The chapter highlights concrete research directions—designing robust contracts and auctions, modeling agent‑driven market dynamics, aligning incentives under delegation, and building governance institutions—which signal both applied opportunities for agent research and urgent social‑policy questions about regulation, liability and organizational design.