Researchers from NTT Research's Physics of Artificial Intelligence Lab and Harvard University's Center for Brain Science have found that adding more AI agents to a task does not necessarily improve results, and beyond a certain point can actually make performance worse. This challenges the "swarm" or multi-agent approach championed by both OpenAI and Anthropic, which implies that running more agents boosts productivity, an assumption that also happens to increase customer spending with those companies.

Using a "Flag Game" test, in which agents must reach consensus on a national flag despite each seeing only a fragment of it, the researchers found the ideal group size was 16 agents. Below that threshold, agents lack enough shared evidence to agree; above it, they tend to split into polarised, competing factions rather than reaching consensus. Elizabeth Pavlova and Hidenori Tanaka, the study's authors, liken this to human organisations, noting that hiring more people does not automatically make a company more effective, and their findings are due to be presented at the ICML 2026 Workshop AI4GOOD. Their advice to enterprise leaders deploying AI agents is to prioritise quality over quantity.