What they did
Will Knight, a WIRED writer, wanted to see whether a swarm of agents could run a Premier League fantasy roster. On a flight he gave Codex a couple of research papers on the topic and asked it to build a harness that checks recent football news and runs a prediction model to suggest a team. When the first picks looked poor, he had Codex split the work across several agents: some read news and blogs, some analysed recent results, some ran models, and others critiqued the suggestions. He then let it pick and manage an experimental team for a few weeks.
How it went
The first single-pipeline version made odd choices, such as benching Erling Haaland. The multi-agent setup gave clearly better suggestions. Raising the count past 100 agents changed little apart from a bigger API bill. After several weeks the team had not scored above the league average, and he said he would not use its picks for his own team.
Worth knowing
Adding more agents did not help; splitting roles (readers, modellers, critics) did. A University of Chicago researcher quoted in the piece says custom harnesses with distinct agent roles work best for forecasting, though top human forecasters still often do better. Expect real API costs if you scale the agent count.