New study reveals how to scale LLM multi-agent AI systems

Researchers from Munich Technical University have published a paper proposing the first systematized framework for scaling LLM-driven multi-agent systems (MAS). The team distilled four core design principles—simplicity, elastic feedback, sequential workflows with optional loops, and summary-based communication—to guide architectural scalability. These principles were implemented in a reference architecture formalized as a constrained directed workflow graph, then evaluated across four increasing complexity configurations using two LLMs of differing capabilities.

The study tested these configurations on standardized terminal-based system engineering tasks. Results showed measurable accuracy improvements with approximately linear cost growth, but crucially, only when the underlying LLM exceeded a minimum capability threshold. Performance peaked at intermediate complexity before degrading due to timeouts and evaluation constraints. Persistent consistency issues also emerged as a key challenge across all scaling levels, highlighting priorities for future research and practical deployment.