I Put a Bunch of AI Agents in Charge of a Fantasy Soccer Team. They’re Not Better Than Humans—Yet

Judgment

Fantasy managers deciding whether to hand team selection to Codex-style agents need a human veto: Codex initially benched Erling Haaland, and scaling past 100 agents mostly raised API cost rather than improving picks.

Basis

Will Knight reports that Codex's first fantasy suggestions were weak—it left Erling Haaland on the bench. A multi-agent harness that read news and blog posts, analyzed recent results, ran models, and critiqued suggestions improved output; raising the agent count above 100 did not help much beyond increasing API cost. Haifeng Xu of the University of Chicago says models are improving because they can consume huge amounts of real-world information and orchestrate complex systems, and he says a custom harness with agents assigned unique tasks is most effective. In his prediction-market tests, GPT-5 ranked best, followed by Grok 4, Claude Sonnet, Gemini, and Llama; Prophet Arena benchmarked forecasting. Prediction Arena gave frontier models $10,000 to wager over 57 days: they consistently lost money but improved over time. Xu notes AI still often trails superforecasters because humans understand real-world dynamics more deeply, though Metaculus Cup evidence suggests that edge may be eroding.