Artificial Intelligence

AI Agents Manage Fantasy Team in Predictive Skill Test

Deploying a swarm of AI agents to manage an experimental roster reveals both the capabilities and current limitations of automated prediction systems.

Radwan AbdulhadiEditor and technology writer

Published: October 29, 2026 · Updated: October 29, 2026 · 3 min read

AI Agents Manage Fantasy Team in Predictive Skill Test — Image: WIRED

Managing a Fantasy Roster With AI Agents

As artificial intelligence continues to expand into various aspects of daily tasks, an experiment was conducted to see if a swarm of bots could successfully manage a Premier League fantasy team .

While traveling, Codex was utilized to build a custom harness consisting of instructions and code to guide different agents through checking recent news and running predictive models. Although the initial setup collated vast amounts of data, it yielded questionable choices, such as leaving Erling Haaland on the bench.

Refining the Multi-Agent Approach

To improve the system, the architecture was expanded into a team of specialized AI agents . Specific agents were assigned to read news and blog posts, analyze recent matches, run predictive models, and critique suggestions.

Scaling the agent count beyond 100 did not drastically improve performance beyond increasing API costs. Nonetheless, the framework highlights a significant technological leap compared to older modeling tools from the UK's Turing Institute, which failed to incorporate real-time world events like player injuries or managerial changes.

Broader Forecasting and Prediction Markets

University of Chicago computer scientist Haifeng Xu studied how commercial models perform on prediction platforms such as Kalshi and Polymarket .

Testing revealed that models could outperform the average person without special guidance. In their assessments, GPT-5 emerged as the strongest forecaster, followed by Grok 4, Claude Sonnet, Gemini, and Llama.

Benchmarks and Evaluating Future-Gazing

To systematically evaluate these forecasting abilities, Xu and colleagues developed Prophet Arena .

Additional evaluation tools include Prediction Arena , a benchmark from Harvard University and Arcada Labs that provided frontier models with funding to wager on prediction markets over a 57-day period.

Although the models initially lost money in that trial, researchers observed consistent improvement over time.

AI Versus Human Superforecasters

Despite rapid advancements, AI still often trails elite human experts known as superforecasters. According to researchers, while humans consume less information, their deeper understanding of real-world dynamics gives them an edge.

Even so, evidence indicates that the human advantage is shrinking. A contest called the Metaculus Cup recently showed that artificial intelligence can occasionally outperform top human predictors, signaling potential shifts for industries rooted in trading and forecasting.