More Agents Makes It Worse, and Washington Wants to Standardize the Ones You Have
Google and MIT Quantify the Multi-Agent Scaling Ceiling
The assumption that adding agents improves performance has been treated as conventional wisdom. A new study from Google Research and MIT, testing 180 configurations across five architectures and three model families, shows it is frequently wrong.
On parallelizable tasks like financial analysis, centralized coordination improved performance by 81% over a single agent. On sequential reasoning tasks, every multi-agent variant degraded performance by 39-70%. The culprit: communication overhead fragments the reasoning process, consuming the cognitive budget needed for actual work. Independent multi-agent systems amplified errors by 17.2x; centralized systems contained amplification to 4.4x. The team also built a predictive model (R² = 0.513) that identifies the optimal architecture for 87% of unseen task configurations based on task properties like tool count and decomposability. Google Research blog post
Before adding agents to your system, measure your task's sequential dependencies and tool density. The paper provides a principled framework for that decision — use it instead of guessing.