Smaller AI agent teams beat larger ones when they remember better

Researchers found that adding more AI agents to a team doesn't automatically make it perform better over time — smaller teams with better memory systems outperform larger ones. This means companies building multi-agent AI systems have a cheaper path to improvement: design better memory for fewer agents instead of just throwing more agents at the problem.

Everyone assumed the scaling rule for AI was simple: more agents equals better results, just like more workers equals more output. This paper shows that assumption breaks down when agents need to learn from their own past work. The practical implication is obvious: if you're building an AI system that has to get smarter through experience, you might spend less money and get better performance by investing in memory design instead of computing more agents. That's a structural insight that changes how companies should architect these systems.