On October 9, Vals AI tested whether teams of coding agents are worth their price. GPT 6 Sol and Claude Opus 5.5 built 50 web apps, each alone and as a team with subagents, at medium and maximum reasoning effort. Teams cost 1.8 to 5.1 times as much. Of the four team-versus-single comparisons, only one was statistically significant: Sol at medium effort, which gained 7.3 points. For Sol, raising the single agent's reasoning effort from medium to max added 11.4 points, more than a team added at medium effort.
That study concerns splitting work, not whether agents make the same mistakes. Still, it raises a question: when you add another model to a vote, how many independent opinions do you actually add? Condorcet's jury theorem from 1785 suggests majority accuracy climbs toward 1 as voters are added, but only if each voter is right more than half the time and their judgments are independent. With p = 0.7, three voters are right 78.4% of the time, five 83.7%, nine 90.1%, and twenty-five 98.3%.
Independence is the fragile assumption. Five friends who studied from the same flawed notes will repeat the same error, so the majority is confidently wrong. Language models trained on overlapping text with similar designs could plausibly share blind spots. A simulation illustrates the gap: with independent mistakes, a five-model majority is right 83.8% of the time, but with overlapping mistakes it is right only 74.4%, even though each model's individual accuracy is unchanged.
The article sets out to measure this effect directly on six real open models across 1,500 exam questions, applying a formula to estimate how many independent opinions a panel truly provides.
Source: Towards Data Science · Summarized by HeadlinesBriefing