More AI agents don’t always mean better results: NTT
CALIFORNIA, UNITED STATES — As enterprises race to deploy agentic artificial intelligence (AI) at scale, new study from NTT Research and Harvard University’s Center for Brain Science finds that multi-agent AI systems perform best within an optimal operating range — beyond which adding more agents actively degrades collective performance.
Peak performance exists, then drops off
The research, conducted using an experimental framework called the Flag Game, tested how groups of AI agents coordinate to solve collective intelligence problems.
The study found that performance peaks at approximately 16 agents before declining as group size grows, a pattern the researchers compare to coordination failures in large human organizations.
Hidenori Tanaka, group leader of the Physics of Artificial Intelligence Lab at NTT Research and of the Physics of Intelligence Program at Harvard University’s Center for Brain Science, said the findings challenge how enterprises think about AI deployment at scale.
“Our research shows that simply adding more AI agents does not necessarily improve performance — just as hiring more people does not automatically make a company more effective,” Tanaka said.
At some threshold, adding more AI agents creates coordination overhead that outweighs any gains in coverage or processing speed.
Design and diversity matter as much as size
“Communication becomes harder, and groups can split into competing camps,” Tanaka said, describing what happens as multi-agent systems grow beyond their optimal range.
The research found that how agents are organized and how they communicate shapes outcomes as much as the number deployed.
Agent diversity also matters: teams combining complementary AI models with different strengths consistently outperformed homogeneous groups of identical agents.
Systems given clear human guidance and communication structures achieved better outcomes than those left to self-organize.
For enterprises building multi-agent AI workflows, the design choices — how agents are structured, how they communicate, and how humans stay in the loop — matter more than headcount alone.
For business process outsourcing (BPO) providers deploying AI agents across customer service, data processing, and back-office operations, the NTT findings carry direct implications.
Scaling agentic AI without governance structures invites the same coordination failures the research documents — competing agent outputs, conflicting interpretations, and degraded throughput.
Teams that invest in agent communication design, role specialization, and human oversight checkpoints are better positioned to deliver consistent, accurate outputs at scale than those that treat deployment volume as the primary metric.
Related news
- EU AI act enforcement begins for AI model makers · 4 Aug
- ClickUp cuts 22% of staff, replaces them with 3,000 AI agents · 27 May
- AI layoffs raise questions over whether automation is driving job cuts · 20 Mar
Disclosure: Outsource Accelerator uses AI tools in the backend of its editorial workflow. Every article is reviewed and verified by a human editor before publication.
Stay ahead of the outsourcing industry. Join thousands of business leaders who rely on Outsource Accelerator for the news, trends, and expert insights that matter. Subscribe to our free newsletter and never miss an update.
