Research by Vals AI indicates that multi-agent systems often fail to outperform single AI agents while significantly increasing operational costs. Testing GPT-6 Sol and Claude Opus 5.5 on the Vibe Code Bench, the study found that team setups cost between 1.8x and 5.1x more than individual agents. Only one out of four tested scenarios showed a statistically significant improvement in performance.
At maximum reasoning effort, team configurations provided no measurable advantage for either model. Anthropic's internal tests with Opus 5.5 showed that increasing the number of agents from ten to 100 resulted in only marginal score improvements after 24 hours. While larger teams can reach performance levels faster, they consistently consume more tokens, leading to diminished returns.
OpenAI researcher Noam Brown noted that multi-agent systems primarily enhance speed rather than output quality. For instance, using four agents can solve tasks twice as fast but doubles the cost. Scaling to very large numbers of agents remains largely unexplored due to prohibitive costs and the coordination tax, where communication overhead between agents degrades overall efficiency.
The effectiveness of agent teams depends heavily on the specific task. While web research and mathematical problems can be parallelized, creative tasks like writing a novel do not benefit from increased agent counts. OpenAI developer Eric Provencher cautioned that agent swarms frequently result in wasted capital as coordination between individual units breaks down during complex operations.
