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

The cost to run AI agent teams can be over five times higher for little to no quality gain.

Research by Vals AI indicates that using teams of AI agents offers minimal quality improvements over single agents, despite significantly higher costs. The company tested GPT-6 Sol and Claude Opus 5.5 on its "Vibe Code Bench" benchmark, comparing solo agents against teams at medium and maximum reasoning efforts. Agent teams were 1.8x to 5.1x more expensive. Only one out of four comparisons showed a statistically significant gain: GPT-6 Sol at medium reasoning, where the team scored 7.3 points higher. At maximum reasoning, no significant advantage was observed for either model using teams. Anthropic's internal tests with Opus 5.5 also showed diminishing quality returns as more agents were added. While larger teams could reach performance targets faster, scaling from ten to 100 agents yielded only slight score increases. OpenAI's research suggests that multi-agent systems primarily offer speed benefits rather than quality enhancements, with coordination costs potentially negating benefits.

The findings challenge the prevailing notion that scaling up AI agent teams directly translates to superior performance, particularly for complex reasoning tasks. The "coordination tax" identified by researchers implies that the overhead of managing multiple agents may outweigh their collective intelligence, leading to diminishing returns on investment. This suggests a need for more efficient agent architectures or task-specific optimizations rather than a brute-force approach to team scaling.

Developers and businesses utilizing AI agent systems should critically re-evaluate their use cases and cost-benefit analyses. The research indicates that current team-based approaches may be economically unviable for many applications, especially when models are already operating at their peak. Future development may focus on optimizing single agent capabilities or exploring novel coordination strategies to mitigate the significant token expenditure associated with agent teams.