During a recent experiment, the author deployed a collection of artificial-intelligence agents to assemble and run a Premier League fantasy football roster. The test was motivated by the rapid spread of autonomous bots into everyday tasks, and aimed to see whether a coordinated swarm could outperform a human manager in a competitive league setting.
An earlier trial using a tool from the United Kingdom’s Turing Institute produced wildly unrealistic line-ups because it ignored current injuries and managerial changes. On a later flight, the author asked the Codex model to assist, first supplying it with relevant research papers and then requesting a code harness that would scrape recent news, analyse match data, and generate player predictions. The first output even suggested benching a top striker, prompting skepticism.
The author then instructed Codex to create a suite of specialized agents: some that read news articles and blogs, others that evaluate recent match results, a set that runs statistical models, and a group that critiques the generated line-ups. Adding more than one hundred agents increased the API bill but did not noticeably improve the quality of the suggested squads.
Computer-science researcher Haifeng Xu of the University of Chicago observes that newer frontier models can ingest massive real-world data streams and coordinate more intricate agent networks, which he believes underlies their growing predictive ability. Xu explains that the key advantage lies in the models’ expanding capacity to understand the world, rather than merely scaling raw parameters.
Xu’s team has been testing AI forecasts on emerging prediction markets such as Kalshi and Polymarket, asking models to guess outcomes ranging from NBA championships to political meetings. Their internal ranking placed GPT-5 at the top, followed by Grok-4, Claude Sonnet, Gemini and Llama. To benchmark performance, they built Prophet Arena, while a separate Harvard-Arcada Labs effort, Prediction Arena, let models wager $10,000 over 57 days; the bots lost money but showed steady improvement.
Despite these advances, AI still trails elite human forecasters, whose nuanced grasp of contextual dynamics remains superior. Nevertheless, contests such as the Metaculus Cup have begun to show AI overtaking top predictors, and upcoming benchmarks like ForecastBench may further narrow the gap. For now, the author’s fantasy squad generated by the agent swarm has not outperformed the league average, and none of its picks have been adopted.