Wired tests a Codex-built AI agent team on Premier League fantasy picks
In his AI Lab newsletter for Wired, Will Knight describes an experiment: a few weeks ago he deployed a small horde of AI agents to build and manage a Premier League fantasy roster. He wanted to know whether a swarm of bots could put together a winning team.
The author had looked at AI for fantasy picks a few years earlier, using a tool from the UK's Turing Institute that chose teams after modelling past fantasy league data. Its suggestions were, in the author's word, bonkers, because it did not take in real-world information such as a player's injury or a team's change in management.
This time, on a flight to visit family, the author asked Codex for help. First came a couple of relevant research papers fed to the model. Then Codex was asked to build a harness, meaning instructions and code that guide different agents through various steps, to check recent news and run a model to make predictions. The setup quickly collated and analysed far more information than the author would have. Even so, the first team suggestions were surprisingly dumb; the author's example was leaving Erling Haaland on the bench.
The author then asked Codex for a team of agents with different jobs: several that read news and blog posts, others that analysed recent results, more that ran models, and a bunch that critiqued the suggestions. This seemed to significantly improve the picks. Raising the number of agents to more than 100 did not change much besides running up the API bill. Still, the author calls the improvement over the Turing Institute tool striking. The final verdict on the results is modest: the AI's team has not outscored the average over the past weeks, and the author will not use its suggestions in their own selections.
The piece then turns to research on AI forecasting. Haifeng Xu, a computer scientist at the University of Chicago who studies AI in market situations, says newer models may predict better because they can consume and analyse huge amounts of real-world information and orchestrate more complex systems. Late last year Xu was watching the growing popularity of prediction markets like Kalshi and Polymarket and, with several colleagues, began testing how the best commercial models forecast. The models guessed, for example, who would win the NBA Championship and who President Trump would meet with. Xu says the models did really well very quickly, and were sometimes able to do better than the average person without special instructions or guidance. In their testing, GPT-5 was best, followed by Grok 4, Claude Sonnet, Gemini and Llama.
Xu's team went on to build Prophet Arena, a benchmark of a model's ability to predict. Another, Prediction Arena, from researchers at Harvard University and a company called Arcada Labs, recently gave frontier models $10,000 to wager on prediction markets over 57 days. The models consistently lost money, but the researchers saw evidence that they improved consistently over time.
Xu says AI predicts most effectively through a custom harness, and that agents with unique tasks, such as reading the news, collecting other data and analysing the whole system, seem to be the most effective approach. AI still often lags behind superforecasters, the very skilled human experts. Xu explains that humans cannot consume as much information, but their understanding of the real world, the dynamics behind a news story, is deeper.
The author argues the human edge is eroding. A contest called the Metaculus Cup recently showed that AI can sometimes beat even the best human predictors, and a benchmark called ForecastBench could add evidence that AI matches human skills when the results of an ongoing contest conclude soon. The author expects the big labs to turn to future-gazing as a way to measure progress in machine intelligence, and suggests that industries built on trading may soon be upended by AI, with the best human traders perhaps supervising and orchestrating AI agents. The author closes by predicting AI will get better, perhaps making fantasy leagues easy to dominate, and says they hope it will not, since it would spoil the fun of winning by yourself.
Key facts
- Will Knight asked Codex to build a harness and a team of agents (news readers, results analysts, model runners, critics) to pick a Premier League fantasy roster.
- Specialised agents seemed to significantly improve the suggestions, but going past 100 agents changed little besides the API bill.
- After a few weeks the AI's team has not outscored the average, and the author will not use its picks.
- Haifeng Xu of the University of Chicago says models did well on prediction markets; in his testing GPT-5 ranked first, followed by Grok 4, Claude Sonnet, Gemini and Llama.
- Prediction Arena gave frontier models $10,000 to wager over 57 days; they consistently lost money but improved over time.
Why it matters
The piece links a small, personal agent experiment to a broader question: whether multi-agent systems can forecast the real world. The author sees striking progress since an earlier Turing Institute tool gave bonkers picks, and cites benchmarks such as Prophet Arena, Prediction Arena, the Metaculus Cup and ForecastBench. The author expects big labs to use forecasting as a way to measure progress in machine intelligence.
Who it affects
Fantasy league players are the immediate audience, but the author's larger point is about industries predicated on trading, which may soon be upended by AI. The author suggests the best human traders may end up supervising and orchestrating AI agents. Superforecasters, the skilled human experts, are said to still hold an edge that is eroding.
How to use it
The pattern described is a harness that guides different agents through steps. The author's setup had agents that read news and blog posts, agents that analysed recent results, agents that ran models, and agents that critiqued suggestions. Xu says agents with unique tasks, such as reading the news, collecting other data and analysing the whole system, seem the most effective approach. Adding agents beyond that, to more than 100, brought little besides a bigger API bill.
How solid is it
This is an anecdotal newsletter column, not a formal evaluation. The fantasy result is a single qualitative statement: the team has not outscored the average over recent weeks. No scores, ranks or API costs are given. The ranking of models comes from Xu's own testing as reported by the author, and no figures are given for Prophet Arena, the Metaculus Cup or ForecastBench. The Prediction Arena losses are not quantified.
Risks and caveats
The author's own team is not beating the average, and its first picks were dumb, such as benching Erling Haaland. The Prediction Arena models consistently lost money, though they improved over time. AI still often lags behind superforecasters. The ForecastBench point is conditional on an ongoing contest concluding soon. Claims about industries being upended are the author's opinion, framed with 'may'.
“Very quickly we started to realize that models did really well”
— Haifeng Xu, computer scientist at the University of Chicago