A new review challenges the common assumption that several communicating language-model agents outperform one model simply because they collaborate. The analysis combines nine empirical evaluations covering 516 distinct multi-agent configurations and compares them through a shared resource-accounting framework.

Many published comparisons put a multi-agent system against one greedy model response. That gives the agent team more samples and more inference compute, making it difficult to tell whether communication caused the improvement. Against baselines matched for sample count or resource budget, the review reports an average accuracy difference of minus 0.042 for multi-agent systems. Accuracy was equal to or below the baseline in 74% of comparable contrasts.

The paper also finds that replacing a single greedy answer with a sampled baseline reduced the measured multi-agent advantage by 0.038 on average. Its statistical sensitivity checks remained negative, though the result synthesizes heterogeneous studies rather than running one controlled benchmark.

The author argues that communication cannot add information when every agent sees the same evidence. Useful gains are more plausible when agents receive different evidence, can verify results externally, or use models with genuinely different error patterns. The paper releases a reviewer checklist and accounting utilities to make future comparisons more explicit about samples, tokens and compute.