AI agents manage fantasy soccer team but fall short of human rivals

AI agents take on fantasy soccer management

WIRED journalist Will Knight deployed over 100 AI agents to manage an experimental Premier League fantasy soccer team, testing their ability to analyze news, results, and predictive models. The AI system was designed to simulate human decision-making by processing real-world data and generating team selections.

Early struggles and incremental improvements

The AI’s initial suggestions were described as impractical—such as benching a top-performing player like Erling Haaland—highlighting gaps in its contextual understanding. However, Knight refined the system by introducing multiple AI agents with distinct roles, including news analysis, results tracking, and predictive modeling. This multi-agent approach improved performance but did not eliminate errors.

Multi-agent systems and rising costs

Scaling the number of AI agents to over 100 yielded marginal improvements in decision-making but significantly increased operational costs, particularly API fees. The experiment underscored the trade-off between complexity and efficiency in AI-driven systems.

AI’s predictive performance: benchmarks and gaps

Research by University of Chicago computer scientist Haifeng Xu found that newer AI models, such as GPT-5, outperform older versions and some humans in predictive tasks. Xu’s benchmarks, like Prophet Arena, evaluate AI’s ability to generate forecasts, while Prediction Arena tested models in real-world prediction markets—where they consistently lost money.

While AI has shown potential in specific contests, such as the Metaculus Cup, it still trails human superforecasters in accuracy and nuance. Ongoing contests like ForecastBench may provide further clarity on whether AI can close this gap.

Fantasy soccer as a real-world test case

The fantasy soccer experiment revealed AI’s strengths in processing large datasets but exposed its limitations in interpreting dynamic, real-world contexts. Despite improvements, the AI-managed team did not outperform the average human competitor in recent weeks, reinforcing the challenges of replicating human intuition and adaptability.

What this means

Lazyfounder analysis — our interpretation, not reported fact.

This experiment matters to founders and operators in AI and predictive analytics because it tests the practical limits of AI agents in dynamic, real-world decision-making. Fantasy soccer—a domain requiring constant adaptation to news, injuries, and performance trends—serves as a microcosm for industries like finance, logistics, or sports management, where predictive AI could automate complex tasks.

The results show that while AI models like GPT-5 are improving in raw predictive power, they still struggle with contextual nuance and cost-efficiency at scale. For startups building AI-driven tools, this signals a need to focus on hybrid systems that combine AI’s data-processing strengths with human oversight, rather than relying solely on fully autonomous agents.

The experiment also highlights the importance of benchmarks like Prophet Arena and ForecastBench. These tools help founders evaluate whether their AI models are ready for real-world deployment—or whether they risk repeating the experiment’s costly missteps, such as over-scaling agents without clear performance gains.

Related on Lazyfounder

• Polymarket
• More ai news
• Prediction markets struggle to escape the gambling label