AI agents can gather more information than a person might manage alone, but that does not guarantee better predictions. An experiment with a Premier League fantasy team found that dividing the work among agents improved the recommendations, yet the team still did not beat the average score.
More agents did not mean a better team
The experiment began with a system designed to check recent news and use a model to make predictions. Its first roster suggestions were surprisingly poor: Erling Haaland was left on the bench.
The setup was then reorganized so agents had separate roles. Some read news and blog posts, others analyzed recent results, some ran models, and others critiqued proposed choices. This approach noticeably improved the suggestions.
Adding more than 100 agents did not make much difference beyond increasing the API bill. The result points to the value of giving agents distinct tasks, rather than simply multiplying their number.
Even after the improvement, the fantasy team did not outscore the average over the weeks described. The writer did not plan to use the AI’s recommendations for their own selections.
Forecasting systems are getting more capable
The fantasy experiment followed an earlier attempt to use a tool from the UK’s Turing Institute, which selected teams by modeling past fantasy league data. That system missed relevant real-world context, such as player injuries and changes in team management.
Haifeng Xu, a computer scientist at the University of Chicago who studies AI in market situations, said newer models may make better predictions because they can process large amounts of real-world information and coordinate more complex systems.
Xu and colleagues tested commercial models on questions such as who would win the NBA Championship and who President Trump would meet with. In their tests, models sometimes did better than the average person without special instructions. GPT-5 performed best, followed by Grok 4, Claude Sonnet, Gemini, and Llama.
Xu described AI as improving in its understanding of the world. That capability can matter when a forecast depends on connecting information from different sources, rather than simply recognizing patterns in past results.
Benchmarks show promise, with limits
Xu’s team developed Prophet Arena to measure a model’s ability to make predictions. Another effort, Prediction Arena, gave frontier models $10,000 to wager on prediction markets over 57 days. The models consistently lost money in that experiment, although researchers saw evidence that their performance improved over time.
These results show why a strong forecasting system needs more than a model that produces a confident answer. A harness can coordinate separate agents to gather news, collect other data, and assess the system as a whole.
AI still often trails highly skilled human forecasters, known as superforecasters. Xu’s explanation is that people may process less information, but bring a deeper grasp of the real-world dynamics behind a news story.
That advantage may be narrowing. The Metaculus Cup showed that AI can sometimes beat even the best human predictors. ForecastBench could offer more evidence about whether AI can match human forecasting skills when results from its ongoing contest conclude.
What better predictions could change
Forecasting experiments raise the possibility that fields built around trading could be reshaped by AI. In that scenario, skilled people might take on a role supervising and coordinating agents that make forecasts.
That possibility remains ahead of the fantasy soccer results. The experimental team was still unremarkable, and its suggestions had not delivered a winning edge. For fantasy players, the appeal of competing and earning bragging rights may depend on keeping some of the decision-making human.
The broader lesson is that AI forecasting is developing through systems that combine information gathering, analysis, and critique. More agents alone did not solve the fantasy team’s problem, and the results described do not show that AI has already surpassed people in every forecasting setting.