Research map / Multi-agent systems and games / LLM agents

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

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces AI-Work, a simulated gig-economy labour market where LLM agents bid for jobs, train skills, and build reputation, and shows that a metacognition/competitive-awareness/planning prompt scaffold (SSA) improves market share over CoT and ReAct.

Problem

Existing frameworks do not explain how AI-only labour markets behave under adverse selection, reputation dynamics, and competitive pressure, nor which reasoning capabilities let LLM agents succeed in such markets.

Approach

A partially observable stochastic game testbed (AI-Work) with latent skills, public Beta-based reputations, CES/Cobb-Douglas job scoring with Gale-Shapley matching, and a bid-or-train meta-action; plus a prompt scaffold (SSA) that elicits metacognition, competitive awareness, and long-horizon planning.

Key results

• SSA achieves 14.3% market share vs 9.7% for CoT and 9.3% for ReAct (about 1.5x) across 25 markets with GPT-5.
• Trace-based capability scores correlate with realized reward (r=0.744 metacognition, 0.699 composite) across 1,600 scored traces.
• Open (revealed-price) bidding induces price deflation (late winning price 0.71 sealed vs 0.61 open) and reduces training; performance-linked pay raises client utility from 0.31 to 0.66.

Surprising findings

• Open winning-price disclosure reduces agent training investment, not only prices.
• Llama 4 Maverick underperforms even the fixed heuristic baselines.

Limitations

• Uses proxy tasks and stylized reduced-form market, not real production work.
• LLM-as-judge trace scoring may introduce measurement error; associations are observational.
• Omits verification, disputes, collusion via communication, and compute/latency costs.

Experimental setup

Datasets: simulated job catalog (AI-Work)

Benchmarks: AI-Work simulated markets

Metrics: market share cumulative reward average reputation training frequency normalised winning bid price win rate Gini coefficient

Baselines: CoT ReAct Specialist Greedy

Builds on: Gale-Shapley stable matching CES utility Beta reputation scheme ReAct chain-of-thought

Closest papers by meaning

• CompeteAI: Understanding the Competition Dynamics of Large Language Model-based Agents (ICML 2024 · similarity 0.88)
• Economy of Minds: Emerging Multi-Agent Intelligence with Economic Interactions (NeurIPS 2026 · similarity 0.87)
• JobBench: Aligning Agent Work With Human Will (NeurIPS 2026 · similarity 0.87)
• Scaling Small Agents Through Strategy Auctions (ICML 2026 · similarity 0.86)
• A Mathematical Framework for AI-Human Integration in Work (ICML 2025 · similarity 0.85)
• TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks (NeurIPS 2025 · similarity 0.85)