Wired’s AI-agent fantasy team still trails skilled humans
🎧 Listen to this story
In an experiment described by Wired, Codex built a Premier League fantasy soccer team and initially suggested benching Erling Haaland. Creating specialized agents for news, recent results, prediction models, and critiques improved the recommendations, but expanding the system to more than 100 agents only increased the API bill, without significant changes.
Haifeng Xu, a computer scientist at the University of Chicago, says newer models may make better predictions by analyzing large amounts of information and coordinating complex systems. In forecasting tests cited by Xu, GPT-5 performed best, followed by Grok 4, Claude Sonnet, Gemini, and Llama. Even so, he says AI often trails so-called superforecasters, whose deeper understanding of real-world dynamics compensates for their lower capacity to process information.
Why it matters · editorial interpretation
The experiment suggests that adding specialized agents does not always improve results proportionally: beyond a certain point, complexity mainly increased the API cost. The comparison with superforecasters also underscores that processing more information does not necessarily replace understanding real-world context.
Sources
ResearchAgents