In short
WIRED tested AI agents on a fantasy soccer team and found that specialized roles improved the system, but it still failed to beat human performance. The story points to real progress in forecasting AI, while showing that human oversight remains essential.
- Specialized AI agents improved fantasy soccer recommendations.
- Adding more than 100 agents raised costs without clear gains.
- The reporter’s AI team still has not outscored the average over recent weeks.
- Research from Haifeng Xu suggests frontier models are getting better at forecasting.
- AI still trails top human superforecasters in many prediction tasks.
WIRED reports that AI agents can now build a fantasy soccer roster that is noticeably better when the system is split into specialist roles, but the setup still falls short of beating humans over time. The reporter’s experiment, run a few weeks ago, showed that more AI coordination helped — yet even a sprawling multi-agent system could not deliver a consistently winning team.
The result matters because it offers a practical snapshot of where agentic AI stands right now: useful, sometimes impressively so, but still unreliable without human oversight and careful orchestration. It also echoes a growing line of research suggesting that newer frontier models are improving not just at language, but at ingesting real-world information and making better predictions.
What WIRED tested
WIRED’s reporter built an experimental fantasy soccer roster using AI agents and compared different ways of asking the system to make decisions. The first version, powered by Codex-generated suggestions, produced poor choices — including benching Erling Haaland, one of the most obvious stars in the player pool.
Fantasy Premier League is a game built around real-world soccer performance, so it is a useful proving ground for AI. Users choose players from actual clubs, then collect points based on how those players perform in matches. That makes the task a mix of sports knowledge, short-term forecasting and decision discipline — all areas where AI often sounds confident before it proves accurate.
To push beyond a single-model setup, the reporter moved to a multi-agent arrangement with distinct jobs: one agent read news, another analyzed results, a third ran model-based comparisons and a fourth critiqued the final choices. The idea was to see whether dividing labor would improve the quality of the lineup.
How the agent setup changed the output
According to WIRED, the split-role system did improve the AI’s recommendations. But adding scale alone did not solve the core problem. Even after the reporter tried adding more than 100 agents, the system did not become meaningfully better — it mainly increased API costs.
That detail is important. It suggests that the value of agents may come less from sheer quantity and more from clear role design, strong prompting and a workflow that forces the model to check its own reasoning. In other words, more bots are not automatically better bots.
Why it matters
The experiment is a small one, but it points to a larger shift in how AI may be used in the real economy. Forecasting, trading, ranking and other judgment-heavy tasks are increasingly attractive targets for AI systems, especially when they can pull in fresh information and break a complex problem into parts.
For readers, the takeaway is straightforward: AI is not yet a plug-and-play replacement for expert decision-making. It can help assemble better inputs, but it still needs human review. For markets and other decision environments, that means the near-term advantage may go to teams that can combine human judgment with well-structured agent workflows, not to systems that merely run at higher volume.
It also matters because fantasy sports are a low-stakes proxy for far more consequential prediction tasks. If a system cannot reliably beat average human performance at choosing soccer players, that sets a caution flag for anyone hoping it can independently manage more complicated forecasting or trading decisions.
What the research side says
WIRED links the reporter’s hands-on experiment to work by Haifeng Xu, a computer scientist at the University of Chicago, whose team has studied how commercial AI models perform in prediction markets. Xu argues that newer models are getting better at absorbing real-world information and coordinating around complex systems.
“AI is just getting smarter in understanding the world”
Xu’s research is relevant because it moves the discussion from one-off demos to structured testing. WIRED reports that his team found strong performance in prediction-market experiments and that GPT-5 ranked best in those forecasting tests. That does not mean AI has solved forecasting, but it does indicate that frontier systems are improving in an area where they once looked weak.
“Very quickly we started to realize that models did really well”
That claim fits a broader pattern emerging across AI evaluation: some frontier models are starting to handle messy, real-world information better than earlier systems. Still, the gap between AI and the best human forecasters has not disappeared. WIRED says AI still trails top human superforecasters in many cases, though that gap may be shrinking.
Background: how we got here
Fantasy sports and forecasting competitions have become useful testing grounds because they combine public data with the pressure of making a prediction before the outcome is known. A few years ago, WIRED says, the same reporter tested an earlier fantasy-team tool from the UK Turing Institute. The new experiment builds on that idea, but with a much more modern agent stack.
At the same time, research groups have been formalizing benchmark-style tests for prediction. Prophet Arena was developed specifically to measure model prediction ability, giving researchers a way to compare systems on forecasting tasks rather than on generic language output alone. Another effort, Harvard/Arcada Labs Prediction Arena, gave models $10,000 to trade over 57 days. The models lost money, but improved over the course of the experiment.
Those tests help explain why the current moment feels different from earlier AI hype cycles. Instead of asking whether a model can generate fluent text, researchers are asking whether it can make useful decisions under uncertainty. That is a harder standard, and the results are mixed.
How prediction testing is evolving
The field is moving toward more realistic evaluation methods. WIRED notes that the Metaculus Cup recently showed AI can sometimes beat top human predictors, and the ForecastBench contest is expected to conclude soon. Together, those efforts suggest a race not just to build bigger models, but to prove that they can reason well enough to perform in public, measurable contests.
For companies building agentic systems, the lesson is clear: prediction is not just a model problem. It is a systems problem. The best results may come from combining retrieval, analysis, critique and domain-specific checks — exactly the sort of division of labor the reporter tried in the fantasy soccer experiment.
By the numbers
The experiment’s key figures show both progress and limitation.
| Metric | What WIRED reported | Why it matters |
|---|---|---|
| a few weeks ago | The reporter deployed the AI agents to manage a fantasy soccer roster | Shows how recent the experiment was |
| more than 100 | Agents were added in an effort to improve the system | Demonstrates that scale alone did not solve the problem |
| $10,000 | Capital used in Harvard/Arcada Labs Prediction Arena | Shows the seriousness of the forecasting test |
| 57 days | Duration of the Prediction Arena experiment | Gives the models time to learn and adapt |
What the fantasy soccer test revealed
The biggest lesson from the reporter’s setup is not that AI failed — it is that AI improved when it was forced into a better workflow. The initial Codex-generated lineup was weak. The specialized-agent setup was better. But even then, the team still has not outscored the average over recent weeks, according to WIRED.
That is a meaningful boundary. It suggests AI can support drafting and lineup selection, but it has not crossed into dependable autonomous management. For now, the system looks best as a decision assistant that needs guardrails, not as a fully trusted manager.
Why specialized agents helped more than more agents
Specialization appears to matter because forecasting is not a single task. One agent can ingest news, another can judge recent form, another can compare model outputs and a fourth can challenge assumptions. That structure can reduce blind spots, but only if the pieces are coordinated well.
By contrast, simply multiplying agents did not create better insight. WIRED reports that more than 100 agents added cost without comparable gains. That is a useful warning for the broader agent ecosystem, where larger numbers of coordinated modules can look impressive in demos while adding complexity faster than they add value.
Human forecasters still have the edge
Even with the progress in prediction benchmarks, humans remain formidable. WIRED says AI still trails top human superforecasters in many cases. That matters because superforecasters are not ordinary hobbyists; they are known for unusually strong performance in prediction tasks, often by combining curiosity, skepticism and disciplined updating.
The current state of the field seems to be this: frontier models are getting better at world understanding and at structured prediction, but the best human forecasters still set the bar. In practice, that means AI can narrow the distance, but not erase it.
What happens next
The next phase to watch is whether the reporter’s AI fantasy team can finally beat the average over a longer stretch of games. WIRED says it still has not, which leaves the experiment open. That ongoing result is useful precisely because it is unresolved.
On the research side, the most important follow-ups are whether specialized multi-agent systems keep improving as models get stronger, whether frontier models consistently beat superforecasters on real-world tasks, and how benchmarks such as Prophet Arena compare with market-based tests. ForecastBench is also nearing its conclusion, which may add another benchmark point to the debate.
If the recent trend continues, the practical use of AI may shift away from “replace the expert” and toward “amplify the expert.” That would still be a major development — especially in environments where speed, synthesis and frequent updating matter — but it would also preserve a central role for human judgment.
Timeline of the story
- A few years ago: the reporter tested an earlier fantasy-team tool from the UK Turing Institute.
- Late last year: Haifeng Xu began studying how commercial models perform on prediction markets.
- A few weeks ago: WIRED’s reporter deployed AI agents to manage a fantasy soccer roster.
- Recently: the Metaculus Cup showed AI can sometimes beat top human predictors.
- Soon: ForecastBench contest results are expected to conclude.
Bottom line
WIRED’s experiment shows that AI agents can get better when they are given specialized roles, but they are still not consistently better than humans at a real-world forecasting task. The progress is real, the limits are real, and the gap between the two is where the most interesting story now sits.
Frequently asked questions
Can AI agents manage a fantasy soccer team better than humans?
Not reliably yet. WIRED’s test found that specialized AI agents produced better lineup suggestions than a basic setup, but the system still has not outscored the average human performance over recent weeks.
Why did the AI fantasy soccer system improve with specialization?
It improved because different agents were assigned different jobs, such as reading news, analyzing results, running models and critiquing decisions. That division of labor appears to work better than asking one system to do everything at once.
What did adding more than 100 agents change?
It mainly increased cost, not performance. WIRED reported that more than 100 agents added little beyond higher API spending, which suggests that structure matters more than raw scale.
How are AI models being tested for forecasting ability?
They are being tested through prediction-market experiments and benchmarks such as Prophet Arena and Harvard/Arcada Labs Prediction Arena. These setups measure how well models predict outcomes under uncertainty, not just how well they write text.
Are AI models catching up to top human forecasters?
They are getting closer, but humans still hold an edge in many cases. WIRED reports that AI still trails top human superforecasters, although the gap may be shrinking as frontier models improve.
