Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableWired→ original

More Than 100 AI Agents Built a Premier League Fantasy Team, and Still Didn't Beat the Average Score

A Wired reporter asked AI agents to build a fantasy football team, where points come from the real matches of Premier League players. Some agents read the news, others analyzed results, a third group ran models, and a fourth criticized the picks. The advice got better. Scaling to 100+ agents only added to the API bill, and the team never beat the average score of other participants. For forecasting, the roles matter more than the number of agents.

AI-processed from Wired; edited by Hamidun News
More Than 100 AI Agents Built a Premier League Fantasy Team, and Still Didn't Beat the Average Score
Source: Wired. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceWired ↗
Editor-in-Chief
Key metric #1GPT-5 was the top forecaster in Prophet Arena, followed by Grok 4
Key metric #2In Prediction Arena, models given $10,000 for 57 days consistently lost money
Key metric #3Superforecasters are still ahead of AI, but in Metaculus Cup the models occasionally won

Fantasy football is a game in which a fan assembles a virtual team from real Premier League players and earns points based on how those people perform in actual matches. You have to guess who will start, who is injured, and whose schedule is favorable, which is essentially making a forecast based on fresh news. Wired columnist Will Knight handed this task to a swarm of AI agents: he had Codex build a system in which several assistant programs read news, run calculations, and argue with each other before selecting a lineup.

More than 100 agents did not produce a win, and the team never beat the average score of other players. For anyone building agentic systems, this is a clear lesson in what actually improves a forecast, and what you are simply paying for on your API bill.

How the Experiment Worked

A few years ago, Knight tried a tool from the Alan Turing Institute that picked lineups based on historical league data. The advice was strange: the system knew nothing about player injuries or coaching changes. This time, on the way to visit relatives, he asked Codex to build a harness, a set of instructions and code that walks the agents through the stages: check fresh news, run a model, and produce a forecast. Before that, Codex was given a couple of scientific papers on the topic.

The first version gathered and analyzed noticeably more information than a human could, but produced absurd decisions. Erling Haaland, for example, was left on the bench. Knight then split the work into roles: some agents read news and blogs, others analyzed recent results, a third group ran statistical models, and a fourth criticized the proposals of the rest. The advice improved.

Next, he decided that more meant smarter and pushed the number of agents to over a hundred. This produced no noticeable effect, but the API bill grew. For an engineer, this is a familiar lesson: if you add identical workers without new sources of information, quality hits a ceiling set by the architecture, while token costs keep climbing.

What Forecasting Researchers Say

University of Chicago professor Haifeng Xu studies the use of AI in markets. In his view, newer models forecast better because they can process enormous volumes of information about the real world and build more complex systems out of several parts. His group tested models on prediction markets such as Kalshi and Polymarket, on questions like who will win the NBA championship or who Donald Trump will meet. Sometimes the models, without special prompting, outperformed the average person.

The Prophet Arena benchmark grew out of this work. In the first tests, GPT-5 performed best, followed by Grok 4, Claude Sonnet, Gemini, and Llama. Xu believes the best results come from a custom harness in which each agent has its own task: reading news, gathering data, and analyzing the overall picture. This matches Knight's experience.

Another benchmark, Prediction Arena, was created by researchers at Harvard and the company Arcada Labs. The models were each given $10,000 to place bets on prediction markets for 57 days. They consistently lost money, although their results improved over time.

Where AI Still Falls Short

The best human experts, so-called superforecasters, are still often ahead. According to Xu, they read less but understand more deeply the dynamics behind a news story. Their edge, however, appears to be shrinking: at the Metaculus Cup tournament, AI occasionally beat even the strongest human forecasters. The results of ForecastBench are expected soon and may provide another comparison. Knight suggests that large labs will start using forecasting of the future as a new way to measure model progress.

There are limitations, too. This is a column, not a controlled study: Knight's team played for several weeks, the sample is small, and the team never exceeded the average score. The author himself does not plan to use the system's advice in his own real-world bets. The Prophet Arena model comparison is an early-stage result, when GPT-5 led, and the ranking could change.

Practical Takeaway

For developers and product managers, the lesson is this. First, forecasting from news requires external data and roles: a fact collector, an analyst, a critic. A single model with a long prompt performs worse. Second, scaling the number of agents must be checked with metrics, not faith: 100+ agents did not justify their cost. Third, a critic that reviews other agents' proposals is an inexpensive component of the design and the first thing worth trying.

Finally, Knight hints at a shift in roles in trading: the best specialists may become managers of AI agents rather than compete with them directly. This is still a hypothesis, but the direction is clear. As for the fantasy league, the author wishes for a bit more human competitive spirit: winning without bot help is still more satisfying.

Story

Key Questions Answered

How the Experiment Worked?

A few years ago, Knight tried a tool from the Alan Turing Institute that picked lineups based on historical league data. The advice was strange: the system knew nothing about player injuries or coaching changes. This time, on the way to visit relatives, he asked Codex to build a harness, a set of instructions and code that walks the agents through the stages: check fresh news, run a model, and produce a forecast. Before that, Codex was given a couple of scientific papers on the topic.

What Forecasting Researchers Say?

University of Chicago professor Haifeng Xu studies the use of AI in markets. In his view, newer models forecast better because they can process enormous volumes of information about the real world and build more complex systems out of several parts. His group tested models on prediction markets such as Kalshi and Polymarket, on questions like who will win the NBA championship or who Donald Trump will meet. Sometimes the models, without special prompting, outperformed the average person.

Where AI Still Falls Short?

The best human experts, so-called superforecasters, are still often ahead. According to Xu, they read less but understand more deeply the dynamics behind a news story. Their edge, however, appears to be shrinking: at the Metaculus Cup tournament, AI occasionally beat even the strongest human forecasters. The results of ForecastBench are expected soon and may provide another comparison. Knight suggests that large labs will start using forecasting of the future as a new way to measure model progress.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…