Experiments deploying swarms of AI agents to manage fantasy soccer teams and participate in prediction markets demonstrate that while frontier models still make basic mistakes, their forecasting ability is steadily improving. Testing by Wired using OpenAI’s Codex to build multi-agent harnesses showed that orchestrating distinct role-based agents—such as news readers, statistical analysts, and critics—substantially outperformed single-prompt setups.

Academic researchers confirm this trajectory. Haifeng Xu of the University of Chicago found that models like GPT-5, Grok 4, Claude Sonnet, Gemini, and Llama performed remarkably well on prediction tasks when evaluating real-world information. Benchmarks like Prophet Arena and Prediction Arena are actively measuring these capabilities, with Harvard and Arcada Labs researchers noting that models wagering on real prediction markets show continuous improvement over time.

However, scaling agent swarms introduces economic trade-offs. Increasing agent counts past a certain threshold yields diminishing intelligence gains while significantly inflating API costs, forcing developers to balance agent architecture complexity against financial utility.

Why it matters

  • Multi-agent architectures outperform single models in complex reasoning, but scaling beyond basic agent swarms rapidly inflates API costs.

  • Improving AI forecasting capabilities will create new automation opportunities in finance, supply chain, and decision-making markets.

Source: wired.com