Why Single Agents Beat Multi-Agent Systems at Equal Token Budgets

A recent study from Stanford found that single-agent systems outperform multi-agent systems in accuracy and compute efficiency when both are given the same token budget. This is due to the Data Processing Inequality, which states that information can only be preserved or lost, not added. Prior benchmarks were biased because they allowed multi-agent systems to spend more tokens than single agents, giving them an unfair advantage. To ensure a fair comparison, researchers should pin the token budget to prevent multi-agent systems from exploiting this advantage.