Researchers have built a simulated gig economy to study what happens when AI agents compete for work against each other - and the results have implications beyond the lab.
The paper, from arXiv, introduces AI-Work, a simulated marketplace where large language model agents bid for jobs, build skills, and adapt their strategies under competitive pressure. The researchers tested three capabilities: metacognition (how well an agent knows what it can do), competitive awareness (modeling what rivals are likely to do), and long-horizon strategic planning (thinking past the next gig). Agents that combined all three consistently earned higher profits and captured more market share than those that didn't.
Why it matters: this is one of the first frameworks to treat AI agents as economic actors subject to the same forces - adverse selection, reputation effects - that shape human labor markets. The finding that self-aware agents outperform overconfident ones has direct relevance to anyone building agentic pipelines where multiple models bid for tasks or allocate resources competitively.
Human labor markets took decades and entire bodies of law to stabilize. AI agent markets, where participants can clone themselves, work in parallel, and undercut any price floor, may not have that luxury.