Markets and institutions in the age of AI Agents
Gillian K. Hadfield and Andrew Koh examine how increasingly autonomous AI systems could transform markets, firms, and economic institutions. Unlike conventional AI tools, these agents can plan and execute complex, long-term tasks with limited human supervision.
The paper’s central message is that economists should not simply treat AI agents as human decision-makers or ordinary software. Although AI agents may appear to optimize objectives like textbook economic agents, their preferences, beliefs, learning processes, and objectives may be unstable or opaque. Humans may therefore be unable to predict what AI agents are actually optimizing.
The authors explore several implications for markets and games:
- AI agents acting as consumers could make purchasing more efficient, but imperfectly specified human preferences could lead agents to buy things that do not reflect what people genuinely want.
- If many agents make similar or biased choices, prices may cease to aggregate information accurately.
- AI agents could reduce search costs and improve matching, but small errors in their representation of humans could produce systematically worse outcomes.
- AI agents may facilitate collusion, strategic bargaining, and new forms of coordination that are difficult for humans or regulators to detect.
- Interactions among AI agents may require new equilibrium concepts because they can condition their behavior on source code, memory, or the anticipated behavior of other agents.
The authors then consider organizations and firms. AI could reduce coordination and monitoring costs, allowing firms to become larger, more diversified, and more concentrated. At the same time, opaque objectives and communication difficulties could make human–AI teamwork harder. The authors also warn that identical or copied AI systems could make errors highly correlated, increasing systemic fragility in finance, supply chains, and other sectors.
A final section addresses the need for new institutional infrastructure. AI agents may require systems for legal identity, registration, reputation records, licensing, monitoring, liability, and contract enforcement. Existing institutions were largely designed for human agents, so legal and regulatory frameworks may need to determine who is responsible when an AI agent acts beyond what its human principal could reasonably foresee or control.
An Economy of AI Agents
Authors: Gillian K. Hadfield, Andrew Koh
From: Johns Hopkins, MIT
Asymmetric understanding and financial stability in the age of agentic AI
Hadfield and Koh, in the above paper, ask what an economy populated by AI agents might look like, while in this paper Markus Brunnermeier asks what happens when humans can no longer understand the agents shaping that economy.
In particular, Brunnermeier argues that humans and AI may not simply possess different information, they may use different representations of the world, making AI decisions difficult for humans to interpret, audit, or anticipate. Brunnermeier calls this asymmetric understanding.
The concept has three central elements:
- AI may understand humans better than humans understand AI. AI systems are trained on extensive records of human behavior and institutions, while their own learned representations and decision rules may remain inaccessible to humans.
- Non-explainability can produce non-alignment. If humans cannot understand how an AI reaches decisions, they may also be unable to determine whether its effective objectives remain aligned with human goals.
- Trust and institutions may be weakened. Finance depends on shared understanding supported by institutions, regulation, professional expertise, and accountability. If AI decisions cannot be independently reconstructed, that institutional trust may erode.
The paper then applies this framework to financial markets. If AI agents become major market participants, prices may become harder to interpret because humans cannot reliably infer the information, strategies, or expectations embedded in AI-generated trades. This could reduce price informativeness, weaken human arbitrage and liquidity provision, increase market instability, and make tacit collusion or manipulation harder to detect.
Central banks could face a particularly difficult strategic asymmetry. AI-equipped market participants may learn the central bank’s reaction function from its speeches, decisions, and historical behavior, while the central bank may be unable to understand the aggregate reaction function of AI-driven markets. This could make financial dominance, policy gaming, and crisis management more serious problems.
As a proposed policy response, Brunnermeier recommends preserving segmented markets with a human or non-AI fallback, relying more on simple and robust rules rather than highly fine-tuned regulation, reconsidering the assumption that greater transparency is always beneficial, and using genuine randomization where predictable policy rules could be exploited. He also calls for stress tests that explicitly include AI concentration, dependency, correlated errors, and loss of human control.
Artificial Intelligence and the Brave New World in Finance
Author: Markus K. Brunnermeier
From: Princeton University
