Researchers have tested whether large language models can function as economic agents in market mechanisms designed for humans, and the results suggest they are not yet ready for the role. A paper published on arxiv in early September found that markets populated by LLM agents converge more slowly, or not at all, toward equilibrium compared to markets with human participants.
What the experiment tested
The study replicated classic economic experiments using a double auction, one of the most widely used market mechanisms. In a double auction, buyers and sellers simultaneously post bids and offers, and trades execute when prices match. Humans in these experiments typically converge quickly toward the competitive equilibrium, the price at which supply meets demand. This convergence is considered a fundamental test of whether a market mechanism works as intended.
The researchers replaced human subjects with LLM agents and ran the same experiments. They checked whether the market delivered an efficient allocation of resources, which is the basic requirement for any market to function. This also tested a dimension of alignment that is rarely discussed: whether LLM agents are compatible with the institutional structures humans have built for economic exchange.
Slower convergence and less efficient outcomes
The core finding was straightforward. Markets with LLM agents exhibited slower convergence toward equilibrium, or failed to converge at all, resulting in less efficient allocations than markets with humans. This means the agents either could not find the right prices, or took much longer to do so, leaving surplus on the table or creating misallocations that human traders would have resolved faster.
The researchers also found substantial heterogeneity across model families and market roles. Some models performed better than others, and the behavior varied depending on whether the agent was acting as a buyer or a seller. This suggests that the problem is not uniform across all LLMs, but the overall pattern was consistent enough to raise questions about using these models as autonomous economic agents.
What Chain-of-Thought traces revealed
The study included a lexical analysis of the Chain-of-Thought reasoning the agents produced during the experiments. This is the internal monologue the models generate to explain their decisions. The researchers found that the decision to execute a trade, rather than continuing to incrementally adjust prices, was associated with a shift in language from strategic reasoning to expressions of urgency.
In other words, the agents were not consistently reasoning their way to trades through deliberate price discovery. Instead, they sometimes jumped to a trade because they felt pressure to act, a pattern that does not align with how humans behave in these settings. The distinction matters because market mechanisms rely on participants making informed, incremental decisions. If agents bypass that process, the mechanism breaks down.
What this means for deploying LLMs in economic roles
The paper does not argue that LLMs cannot participate in markets at all. It argues that the mechanisms humans use to make markets work are not yet compatible with how these models reason. A double auction is a simple mechanism, one of the most studied in economics. If LLM agents struggle with it, more complex mechanisms like procurement auctions or sequential bargaining protocols will be harder still.
For teams building systems where LLM agents make purchasing decisions, allocate resources, or negotiate with other agents, the results suggest caution. The models may produce outputs that look like market behavior, but they do not converge on efficient outcomes the way humans do. This is a form of misalignment that is easy to miss if you are only looking at whether the agent produces a plausible-looking action rather than whether the overall system achieves the intended result.
The researchers have released their testing framework publicly, which should make it easier for others to evaluate additional models or market mechanisms as they become available. The work adds to a growing body of evidence that using LLMs as economic agents requires more than prompting them to behave like humans. The institutional context matters, and the models are not yet adapted to it.