I Put a Bunch of AI Agents in Charge of a Fantasy Soccer Team. They’re Not Better Than Humans—Yet
AI-assisted digest, checked against the original source before publishing. How we work
Codex-run agents assembled a fantasy soccer roster, whiffed on Erling Haaland, then improved—while research suggests frontier models already forecast better than the average person.
A few weeks ago, the author handed an experimental Premier League fantasy squad to a swarm of AI agents he put together with Codex on a flight to visit family. Given how far agents have spread into everyday life, the test was simple: could a horde of bots build a winning team? The first lineup came back poor enough to bench Manchester City striker Erling Haaland.
Codex produced a harness—instructions and code guiding agents through steps such as checking recent news and running a prediction model. Work was then split among specialists: readers of news and blog posts, analysts of recent results, model runners, and critics that poke holes in each other's proposals. That division notably sharpened the suggestions. Scaling the swarm past 100 agents, in hopes of extra intelligence, changed little besides the API bill.
The gain stood out against an earlier attempt: a few years ago, a tool from the UK's Alan Turing Institute selected teams by modeling previous fantasy-league data, and its suggestions were, in the author's words, bonkers, since it never folded in real-world facts like injuries or a change of manager. Haifeng Xu, a University of Chicago computer scientist who studies AI in market situations, offers a wider read on the trend.
‘AI is just getting smarter in understanding the world—I feel like that's the fundamental reason they're improving,’ Xu tells WIRED. Late last year he was struck by the growing popularity of prediction markets such as Kalshi and Polymarket, so alongside several colleagues he began testing how the best commercial models would fare on them. ‘Very quickly we started to realize that models did really well,’ he says, noting they sometimes outperformed the average person without any special instructions or guidance.
Plenty stays unsettled. It is not obvious whether the improvement comes from having more agents or better-orchestrated ones, and whether forecasting skill on political and sports markets carries over to fantasy-style roster decisions. Also worth tracking: how Xu's model rankings move as new releases land, whether the added machinery pays back its API costs, and how much the author's own AI-powered crystal ball—which forecasts rapid gains in frontier models' predictive ability—gets right.
Key facts
- • Xu's team had models predict the NBA Championship winner and Trump's meetings.
- • GPT-5 topped Xu's forecasting tests, ahead of Grok 4, Claude Sonnet, Gemini, Llama.
- • Codex was fed two relevant research papers before building the agent harness.
- • Agents have not yet outperformed humans managing fantasy teams, the experiment concludes.
Why it matters
For builders, the ceiling on agent swarms appears to be about structure, not headcount: splitting agents into reader, analyst, and critic roles produced gains that raw parallelism did not. Expect agent frameworks to compete on orchestration quality and cost-per-insight rather than scale, and watch whether prediction-market accuracy actually converts into better operational decisions.
Spotted an error? Send a correction.
Get the day's AI news in one email
The Daily Epoch · 60-second read · unsubscribe anytime
