Problem

The paper addresses a critical gap in understanding AI labor markets. Current agentic research often overlooks key economic forces like adverse selection, moral hazard, and reputation dynamics, which are fundamental to real-world labor markets. These forces arise due to incomplete information and imperfect monitoring, requiring advanced strategic thinking and self-awareness from AI agents. Existing frameworks lack the capabilities to model how AI agents make autonomous labor decisions, affect long-term profits, and influence economic structures, especially under competitive pressure.

Method

The authors propose a novel framework, modeling the AI labor market as a "Competitive Skill-Based Stochastic Game." This framework is implemented in "AI Work," a simulated gig economy platform that incorporates job allocation, skill development through training and on-the-job learning, and reputation building. Agents' strategic actions include skill development and competitive bidding for jobs, with decisions influenced by latent skills and public reputation. The framework captures adverse selection (unobservable abilities), moral hazard (unobservable effort), and reputation systems (information storage). Experiments involve both fixed-policy agents for macroeconomic analysis and LLM agents (e.g., GPT-5, Llama-4) to study strategic behavior. They identify and quantify three core strategic capabilities in successful LLM agents: metacognition (self-assessment), competitive awareness (rival modeling), and long-horizon strategic planning. A key part of the method is the development of "Strategic Self-Improving Agents" (SSA), which are explicitly prompted to reason across these three domains, and then compared against standard LLM agents (Chain of Thought, ReAct).

Results

Simulations with fixed-policy agents reproduced classic macroeconomic phenomena, such as the inverse hyperbolic relationship between unemployment and job vacancy rates (Beveridge Curve) and the linear relationship between changes in unemployment and aggregate output (Okun’s Law), indicating the model's fidelity. With LLM agents, most performed better than static policies, with GPT models generally excelling. The research revealed that open bidding mechanisms led to aggressive undercutting and systemic wage deflation, discouraging long-term skill investment. Conversely, flat-fee contracts reduced training incentives, lowering overall market utility. Strategic Self-Improving Agents (SSAs), explicitly prompted for metacognition, competitive awareness, and strategic planning, consistently demonstrated superior performance, achieving higher cumulative rewards, better average rank, and larger market share compared to standard LLM agents (CoT, ReAct). Ablation studies showed that metacognition was the most significant driver of economic performance, enabling better specialization and disciplined bidding, while explicit planning prompts had limited additional effect. SSAs also exhibited superior adaptability to market shifts, such as changes in demand or recessionary periods.

Implications

The findings highlight that market design choices significantly influence AI labor market dynamics, impacting wages, investment in skills, and wealth concentration. For instance, sealed bidding and performance-based pay can encourage skill investment and client utility, while open bidding can lead to deflationary spirals. The inherent characteristics of AI agents, like concurrency and replicability, can amplify market inequality, though job diversity can partially mitigate this by encouraging specialization. The study underscores the importance of developing advanced strategic capabilities in AI agents, particularly metacognition, for successful navigation of competitive labor markets. Future research should explore the interplay between reputation systems and explicit verification mechanisms, as well as consider more complex market factors, to fully understand and guide the development of AI-driven economies. The work also has implications for human labor, suggesting potential for rapid monopolization by AI agents and the necessity for human reskilling to adapt to evolving job markets.