I Put AI Agents in Charge of a Fantasy Soccer Team—Here’s How They Compared to Human Managers

To test whether AI agents could make better decisions than a human, I asked Codex to build and manage a Premier League fantasy soccer team. The experiment showed promise—but the team has yet to outperform an average player.

A few years ago, I tried a tool from the UK’s Turing Institute that chose fantasy lineups using past league data. Its recommendations were often wildly off because it missed important real-world context, such as injuries and changes in team management.

This time, I gave Codex research papers and asked it to create a system that could check current news and use models to predict player performance. The first recommendations still had obvious flaws: the system wanted to bench Erling Haaland.

I then asked Codex to divide the work among specialized agents. Some read news and blogs, others analyzed recent results or ran predictive models, while another group critiqued the recommendations. This improved the suggestions. Increasing the swarm to more than 100 agents, however, made little difference—apart from increasing the API bill.

The broader progress in AI forecasting may reflect models’ growing ability to process real-world information and coordinate more complex systems. Haifeng Xu, a computer scientist at the University of Chicago who studies AI in markets, says models can already make predictions that outperform the average person, even without special instructions. In Xu’s tests, GPT-5 performed best, followed by Grok 4, Claude Sonnet, Gemini, and Llama.

Xu and colleagues created Prophet Arena to evaluate AI forecasting. Another study, Prediction Arena, gave frontier models $10,000 to wager on prediction markets for 57 days. The models lost money overall, though researchers saw their performance improve over time.

Xu says the strongest approach is to give agents distinct jobs—such as gathering news, collecting other data, and analyzing the results—within a custom system. Even so, AI often trails skilled human forecasters. People may process less information, but they can better understand the real-world dynamics behind events.

That human advantage may be narrowing. The Metaculus Cup has shown that AI can sometimes beat top human predictors, and ForecastBench may offer further evidence as its ongoing contest concludes. Major AI labs could increasingly use forecasting as a way to measure progress.

For now, my experimental team has not outscored the average, and I won’t be relying on its picks. But prediction tools are improving, and AI could eventually reshape fields built around trading and forecasting—with human experts shifting toward supervising and coordinating agents.

That might make fantasy leagues easier to win. I’m not sure I want it to: part of the fun is earning the bragging rights yourself.

Total
0
Shares
Leave a Reply

Your email address will not be published. Required fields are marked *

Previous Article

ICE Is Conducting One of the Largest Anti-Left Surveillance Operations in U.S. History

Next Article

Oracle Adds Google TPUs to OCI, Expanding Enterprise AI Compute Options

Related Posts