Competitive Market Behavior of LLMs

URL SCAN: Competitive Market Behavior of LLMs
FIRST LINE: # Computer Science > Multiagent Systems

The Dissection

This paper tests whether unmodified LLM agents can operate efficiently inside a human-designed double-auction market. It finds that they converge more slowly or fail to converge, allocate resources less efficiently, vary substantially by model and role, and shift from strategic reasoning to urgency when executing trades.

The paper is therefore measuring protocol compatibility, not the terminal viability of AI capitalism. It examines whether current conversational agents can imitate human market participants under a narrow experimental setup. It does not measure AI productivity, automation of labor, ownership of compute, tool use, persistent memory, delegation, software integration, or output per unit cost.

The Core Fallacy

The central error is treating poor performance in one human market mechanism as evidence against AI’s broader economic dominance. That inference does not follow.

A raw LLM failing to clear a double auction efficiently is an implementation defect, not a structural refutation of the Discontinuity Thesis. The thesis does not require every model instance to be a competent trader. It requires AI systems, once integrated with tools, memory, optimization loops, software agents, and institutional control, to outperform human labor across economically relevant cognitive tasks.

Indeed, the result may expose a transition tax rather than a survival mechanism. If LLM agents initially distort markets, owners can redesign the protocol, add automated pricing layers, constrain agent behavior, or remove humans and language models from direct participation altogether. The market interface can be replaced. Human employment cannot be restored by preserving an inefficient auction ritual.

Hidden Assumptions

• That the double auction is a sufficiently representative test of economic agency rather than one legacy mechanism among many.
• That current LLM agents, operating without richer memory, tools, learning, or institutional scaffolding, represent mature AI capital.
• That lexical analysis of Chain-of-Thought traces reliably reveals the causal basis of trading decisions.
• That slower equilibrium convergence implies inferior long-run economic performance rather than temporary deployment friction.
• That market efficiency under human-designed rules is the relevant alignment criterion for future AI-mediated economies.
• That model heterogeneity will remain economically decisive after systems are selected, fine-tuned, orchestrated, and embedded in competitive production networks.
• That preserving existing market mechanisms matters more than replacing them with mechanisms optimized for machine participants.

Social Function

Classification: partial truth and transition management.

The paper punctures the childish claim that LLMs are already rational economic actors. That part is useful. It documents a real coordination weakness: present agents can be erratic, impatient, role-sensitive, and inefficient under strategic pressure.

But its framing quietly narrows the battlefield. It risks converting “today’s agents are bad traders” into “AI cannot economically displace humans.” That is ideological anesthesia. The paper studies whether machines fit an inherited human interface; the Discontinuity Thesis asks whether machine-controlled production eventually makes most human labor unnecessary. Those are different questions.

The Verdict

This is evidence of lag, not reversal. Current LLM agents are poor inhabitants of a human auction house, but that does not protect the human wage-consumption circuit. The likely outcome is not the restoration of productive participation; it is protocol redesign, concentrated ownership, and machine-managed allocation. The paper identifies a brittle deployment layer while leaving the structural kill mechanism untouched.