AI agent teams waste massive tokens for barely measurable quality gains, research finds

What changed

New research from Vals AI shows that assembling teams of multiple AI agents only slightly improves quality compared to using a single AI agent. The study found that using agent teams consumes up to five times more tokens but yields almost no measurable performance gains. In tests with GPT-6 Sol and Claude Opus 5.5 models, only one out of four team configurations delivered a noticeable improvement. Anthropic’s own findings reinforce this, showing that adding more than ten agents leads to a plateau in quality while costs keep rising.

Why builders should care

For developers and operators relying on AI agents, this means that scaling up with agent teams is inefficient and expensive. Token usage is a direct cost factor with commercial AI APIs, so burning 5.1 times more tokens for negligible boosts hits budgets hard without clear ROI. It also adds complexity to system design with coordination overhead but without clear improvements. This undercuts the expectation that more agents automatically translate into better results.

The practical takeaway

Focus efforts on optimizing single agents or small, well-structured teams rather than trying to multiply agents indiscriminately. Experiment carefully to identify if any multi-agent collaboration truly adds value for specific tasks. Keep a close watch on token consumption and prioritize models or configurations that balance cost and incremental gain. Discard strategies that rely on arbitrarily larger teams as a shortcut to quality because they are likely to waste resources.

What to watch next

Monitor further third-party benchmarks and real-world case studies that explore the scaling of agent teams in diverse applications. Pay attention to responses from AI platform providers and any technological advances that may reduce overhead or improve collaboration efficiency between agents. Stay alert for alternative architectures or methods that could deliver better quality gains without ramping up token costs as dramatically.

AI Quick Briefs Editorial Desk