Multi-Agent AI Systems Often Costlier and Less Accurate Than Single-Agent Setups

Research suggests that deploying multiple AI agents in coordination does not reliably improve reasoning and may actually degrade it. Google's research found that multi-agent coordination reduced sequential reasoning performance by 39 to 70 percent compared to single-agent approaches. A Sber whitepaper also reported that multi-agent setups consumed roughly 15 times more tokens than single-agent chat. Experts argue that each agent-to-agent handoff acts as a lossy step, with downstream agents receiving only sanitised summaries rather than full reasoning context. Multi-agent architecture is considered justified only when tasks are genuinely parallel or exceed a single model's context window.