A recent study by Anthropic challenges the assumption that more AI agents always lead to better outcomes. The research shows that while independent tasks like vulnerability scanning benefit from parallel agents, scenarios involving shared resources, dynamic dependencies, or conflicting goals suffer from increased coordination costs. These costs manifest as duplicated work, collective misjudgment, and behavioral convergence, where agents lose diversity and make similar errors. The findings suggest that system architects should evaluate task decomposability and inter-agent dependencies before scaling agent counts. For tasks that require tight collaboration, investing in better communication protocols and conflict resolution mechanisms may yield higher returns than simply adding more agents. This research provides a practical framework for designing multi-agent systems that balance coverage and coordination efficiency.
Anthropic's research reveals that adding more AI agents can reduce performance due to coordination overhead and conflicts. The key is designing for effective collaboration, not just scaling agent count.