An AI experiment with fantasy soccer

The author deployed AI agents to build and manage a Premier League fantasy team to test whether a swarm of bots could assemble a winning roster.

Earlier attempts were unrealistic

A few years ago, a tool from the UK's Turing Institute selected teams by modeling fantasy league data but ignored real-world information like player injuries or management changes.

Codex built a multi-agent system

The author asked Codex to create agents that read news, analyzed results, ran prediction models, and critiqued suggestions. This approach significantly improved team recommendations.

More agents didn't help much

Increasing the number of agents to more than one hundred did not substantially improve suggestions, only raising the API bill without meaningful gains.

Frontier models show real promise

According to Haifeng Xu at the University of Chicago, newer models may excel at predictions because they consume and analyze huge amounts of real-world information and orchestrate complex systems.

“AI is just getting smarter in understanding the world, I feel like that's the fundamental reason they're improving.”

Testing models on prediction markets

Xu and colleagues tested frontier models on prediction markets like Kalshi and Polymarket. GPT-5 performed best, followed by Grok 4, Claude Sonnet, Gemini, and Llama.

“Very quickly we started to realize that models did really well.”

Humans still have an advantage

AI often lags behind superforecasters because humans understand real-world dynamics more deeply, though evidence suggests this edge is eroding as AI improves.

The fantasy team underperformed

The author's AI-managed team has not outscored the average over recent weeks and will not influence the author's own selections, but AI will likely improve rapidly.