The Agentic Sprawl: Why More AI Bots Don't Equal More Productivity
AI-generated image · US National Wire
As OpenAI and Anthropic push multi-agent architectures, new research suggests that 'agent armies' can lead to polarization and operational inefficiency.
For the modern enterprise, the allure of the 'agent army' is strong. Both OpenAI and Anthropic have championed the deployment of multi-agent systems—OpenAI previously utilizing the term 'swarm' before introducing the OpenAI Agents SDK, and Anthropic focusing on multi-agent orchestration—under the implicit premise that increasing the number of agents scales productivity.
However, from an operational lens, this drive toward quantity risks creating a new form of technical debt. As The Register reported, researchers from Harvard University's Center for Brain Science (CBS) and NTT Research's Physics of Artificial Intelligence (PAI) Lab warn that there is a critical tipping point where adding more agents actually hinders performance.
In work scheduled for the ICML 2026 Workshop AI4GOOD, Pavlova, a senior data scientist at Harvard CBS, and Tanaka, head of NTT's PAI Lab and associate professor at Harvard CBS, describe a "Flag Game" they created to test how agents balance private evidence against social input. In this simulation, agents are given fractional views of a national flag and must communicate—via broadcasting, passing guesses to listeners, or interacting with a manager system—to reach a consensus.
***Opinion:*** *The 'Flag Game' serves as a potent metaphor for enterprise SaaS deployments. When organizations deploy disparate bots without a lean, governed orchestration layer, they aren't building a collaborative workforce; they are building silos that compete for the 'truth' of the data.*
Pavlova and Tanaka found that the ideal agent count for the Flag Game is 16. With fewer participants, the system cannot collect enough evidence to reach a consensus. Exceed that number, and the agents polarize into opposing camps. This suggests that rather than converging on a solution, a bloated agent workforce begins to validate its own collective positions within fragmented groups.
Hidenori Tanaka noted in a statement that adding more AI agents does not automatically improve performance, comparing the phenomenon to hiring more people without a corresponding increase in effectiveness. He warned that as the agent population grows, communication becomes more difficult and groups can split into competing factions.
For enterprise leaders, the takeaway is clear: the goal should not be the largest possible agent workforce, but the most efficient one. To avoid operational chaos, the pivot must move away from raw quantity and toward quality and governed orchestration.