Author uses Codex to build and test AI agents for managing a fantasy team
The author of an article for Wired used Codex to build an agent system for fantasy football. According to him, assigning separate roles improved the suggestions, while increasing the number beyond 100 mainly raised the API bill. The team did not beat the average during the weeks tracked.
The author of an article for Wired supplied Codex with several research papers and had it create instructions and code for an agent system intended to monitor current news and generate fantasy football predictions. After weak initial suggestions, he divided up the work: some agents read news and blog posts, while others analyzed recent results, ran predictive models or critically reviewed suggestions from other agents.
According to the author, assigning separate roles significantly improved the suggestions. However, increasing the number of agents to more than 100 brought no further significant change and increased the API bill. The experimental team did not beat the average during the weeks tracked, and the author said he would not use the suggestions for his own player selections. This reflects his experience with one system, rather than evidence that using more agents generally leads to improvements.
Why it matters
The case offers a concrete architecture for a task that combines current news, results analysis, predictions and a review of suggestions. It also illustrates a reason to evaluate recommendation quality and costs separately: in this experiment, a larger number of agents increased spending without the author observing a substantial improvement.
Two audiences, two different impacts
What this means
For individuals
For fantasy football players, the experiment has yet to demonstrate an advantage in lineup selection: even after adjustments to the system, the team did not beat the average during the weeks tracked.
For a business
For teams developing agent systems, this case provides indirect evidence to inform decisions about API costs: according to the author, increasing the number of agents beyond 100 did not produce a corresponding improvement in suggestions in this experiment.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.