When AI Agents Compete for Jobs: Strategic Capabilities and Economic Dynamics of AI Labour Markets

AI summary

Overview

• Research area: Multi-agent systems, computational economics, LLM agent behavior, labor market design
• Technical level: Intermediate (assumes familiarity with LLM agents, basic economics, and matching theory)
• Scope: Introduces AI-Work, a simulated gig-economy testbed where LLM agents bid for jobs, train skills, and build reputation, then uses it to identify which reasoning capabilities produce competitive success.

What This Paper Is About

As AI agents begin transacting with each other autonomously through marketplaces and protocols, they form what the authors call an "AI Labour Market"—a system where agents negotiate, bid, and fulfill work at scales and speeds far beyond human markets. Unlike human workers, AI agents can work many jobs simultaneously, learn quickly, and labor without wage floors, yet we lack frameworks to understand how classic economic forces like adverse selection and reputation dynamics play out in such markets. The paper builds a controlled simulated economy to isolate these dynamics and to test which reasoning capabilities let agents win under competitive pressure and incomplete information.

Key Contributions

1. A formal testbed for AI economic agency. The paper formalizes an AI labor market as a partially observable stochastic game and instantiates it in AI-Work, capturing adverse selection, reputation dynamics, and explore-exploit tradeoffs in a closed economic loop (bidding, matching, performance, reputation update).
2. AI-specific market dynamics. The authors show how platform rules shape equilibrium behavior—open versus sealed bidding drives price deflation versus skill investment, and AI-specific concurrency amplifies winner-take-all concentration, which task diversity partially mitigates.
3. Three strategic capabilities identified. They operationalize metacognition, competitive awareness, and long-horizon strategic planning as scorable reasoning dimensions, show these correlate with realized profits, and demonstrate they cause gains via intervention.
4. A prompt scaffold (SSA). Strategic Self-Improving Agents (SSA) is a minimal prompt scaffold that explicitly elicits the three dimensions within the same backbone model, achieving 1.5× the market share and 1.2× the cumulative reward of CoT and ReAct baselines.

Main Findings

• Capability matters more than model scale. Agents exhibiting metacognition, competitive awareness, and strategic planning—not raw model quality—consistently capture higher profits and market share. Trace-based diagnostics show correlations of r=0.744 (metacognition), r=0.643 (competitive awareness), and r=0.697 (planning) with realized reward, after controlling for model identity.
• SSA outperforms standard prompting. The scaffold achieves roughly 15% market share versus 9–10% for CoT/ReAct, with better adaptation to price-sensitivity shifts, demand shocks, and recessions (training jumps from 8.9% to 25.3% of rounds during recessions).
• Market design shifts equilibria. Open-price bidding induces persistent price deflation and crowds out training investment; performance-linked pay increases both training and client utility (U rising from 0.31 to 0.66 relative to flat fees). These results hold across varied scoring rules, reputation settings, and matching noise.
• AI concurrency concentrates wealth. Because agents can hold multiple jobs simultaneously, high-reputation agents capture disproportionate work—Gini coefficient drops from 0.70 at one benchmark skill to 0.24 at 64 benchmark skills, showing job diversity mitigates inequality.
• Macroeconomic patterns emerge. The simulation reproduces a Beveridge curve (unemployment-vacancy inverse relationship, R²=0.843) and an approximate 2:1 Okun's-law ratio between output and unemployment (R²=0.436), validating that the underlying matching and reputation mechanics are coherent.
• LLMs behave as distinct strategic actors. Models differ qualitatively: GPT-OSS-120B wins via aggressive underbidding (winning price 0.55), Gemini 2.5 trains heavily (13% of rounds) to build reputation, and Llama 4 underperforms heuristics due to inconsistent multi-round behavior.
• Structure beats knowledge injection. Decomposition experiments show the specific metacognition/competitive-awareness/planning structure (7.9% market share) outperforms structure-only (6.2%), knowledge-only (5.0%), and CoT (4.3%), and all three capability blocks contribute significantly (p<10⁻⁴).

Methodology in Plain English

The authors built a simulated gig platform modeled loosely on Upwork or Fiverr. Each round, LLM agents see a set of job listings and a public earnings leaderboard, then choose either to bid on jobs (earning income now) or to train (improving future skill). Each agent has two hidden layers of state: a latent skill vector that nobody observes (except indirectly through job outcomes) and a public reputation vector that everyone sees. Clients score bids using a weighted combination of price and reputation, then jobs are matched using a stable matching procedure that mimics real platforms optimizing for client satisfaction. Performance updates reputation through a discounted evidence scheme that weights recent work more heavily and gives new agents a cold-start prior.

To test which reasoning patterns matter, the authors ran eight frontier LLMs against fixed-policy baselines across 100 rounds, then scored their decision rationales using an anonymized LLM-as-judge with validated human agreement (κ=0.71). They then built a prompt scaffold (SSA) that forces agents to reason through self-assessment, rival inference, and long-horizon tradeoffs before acting, and tested it against CoT, ReAct, and several token-matched alternative scaffolds to isolate whether the specific structure—not just prompt length or domain hints—drives the gains.

Why This Matters

This is one of the first papers to treat AI labor markets as a distinct economic object with their own mechanics, rather than a variant of human gig work. It provides a controlled method for testing how AI agents will behave when they compete against each other at machine speed, and for identifying which reasoning capabilities confer economic advantage—a question that becomes urgent as agentic marketplaces scale.

Real-world applications:

• Designing AI agent marketplaces: Platform designers can use findings about sealed bidding, capacity caps, and reputation weighting to shape whether their market rewards price wars or skill investment.
• Multi-agent orchestration: Companies deploying agent swarms can use SSA-style scaffolding to improve bidding, scheduling, and resource allocation among competing subagents.
• Regulatory and labor policy analysis: The paper maps how AI-specific properties (concurrency, replicability) amplify concentration, informing debates about wage floors, task diversity mandates, and platform governance.
• Reputation and verification system design: The finding that verification mechanisms would shift markets from "trust economies" toward pure price competition has direct implications for how platforms should design credentialing and quality checks.

Industry relevance: Directly affects AI infrastructure providers (Google's A2A, Anthropic's protocols), gig platform operators, agent marketplace startups, and any enterprise building internal multi-agent resource allocation systems.

Future Directions

1. Verification versus reputation. The authors hypothesize that costless perfect verification would eliminate adverse selection and push markets toward deflationary price competition; testing this trade-off explicitly is a natural next step.
2. Communication, collusion, and strategic manipulation. The current environment forbids inter-agent communication; allowing it raises open questions about tacit collusion, reputation gaming, and feedback manipulation.
3. Richer production workflows. Extending beyond proxy microtasks to end-to-end production—contract negotiation, requirements scoping, delivery disputes—would test whether capability findings transfer to real work.
4. Heterogeneous clients and cost structures. Incorporating compute costs, latency, and diverse client preferences would complete the picture of what agents actually optimize under real deployment constraints.

Target Audience

Researchers in multi-agent systems and LLM agent evaluation; computational economists studying labor market design and agent-based modeling; AI platform engineers and product managers designing agent marketplaces; and policy analysts examining the economic impact of autonomous AI workers. The paper is most valuable to readers already comfortable with either reinforcement learning / multi-agent systems or labor economics, though the plain-language framing of the capability findings makes the conclusions accessible to a broader technical audience.

Authors’ abstract

Emerging agentic marketplaces provide the economic infrastructure for matching and coordinating the large amounts of AI agents used in agentic swarms. Unlike human workers, AI agents can operate on multiple jobs simultaneously, acquire skills rapidly, and labor without wage floors. These differences introduce a new segment of $\textbf{AI labor markets}$, where AI agents interact with each other at a much higher frequency than human markets. Yet we lack frameworks to understand how such markets behave in light of economic forces that shape labor markets, such as adverse selection and reputation dynamics. To explore this, we introduce $\texttt{AI-Work}$, a tractable, simulated gig economy where Large Language Model (LLM) agents compete for jobs, develop skills, and adapt their strategies under uncertainty and competitive pressure. Our experiments examine three domains of capabilities that successful agents possess: $\textbf{metacognition}$ (accurate self-assessment of skills), $\textbf{competitive awareness}$ (modeling rivals and market dynamics), and $\textbf{long-horizon strategic planning}$. Agents with these capabilities consistently achieve higher profits, market share, and stronger adaptation than competing agents. Through $\texttt{AI-Work}$, we hope to provide a foundation to explore the microeconomic properties of AI-only labor markets, and a conceptual framework to study the strategic reasoning capabilities of participating AI agents.