Testing by evaluation firm Vals AI shows that multi-agent teams deliver negligible quality improvements over single AI agents while dramatically inflating operational costs. Evaluating GPT-6 Sol and Claude Opus 5.5 on the Vibe Code Bench, researchers found that agent teams increased token expenses by 1.8x to 5.1x. Out of four test scenarios across medium and maximum reasoning efforts, only one showed a statistically significant performance increase.
The findings align with internal testing at Anthropic and statements from industry researchers. Scaling multi-agent setups from 10 to 100 agents yielded diminishing returns, offering slight speed improvements on specific tasks like math and web search, but failing to improve non-parallelizable workflows. In some cases, performance declined as team sizes expanded.
Industry leaders warn that coordination overhead creates severe inefficiencies in multi-agent architectures. OpenAI researchers noted that multi-agent systems primarily purchase speed rather than higher quality, cautioning developers against deploying large agent swarms due to high costs and degraded coordination.
Why it matters
Avoid multi-agent swarms for complex quality-driven tasks; they multiply API expenses with minimal accuracy gains.
Use multi-agent setups selectively for time-sensitive, highly parallelizable tasks like web search rather than general workflows.
Source: the-decoder.com



