Multi-Agent AI Setup Matched Single Agent on Every Metric in Controlled Test

A software developer ran a controlled experiment comparing a single LLM-powered support agent against a multi-agent team handling triage, refunds, and knowledge retrieval. Both architectures were tested against identical evaluation scenarios measuring safety, intent accuracy, groundedness, and other properties. The results showed zero difference across all five metrics, with not a single scenario changing its verdict. The multi-agent version added complexity — five types instead of one, 127 lines of code versus 91, and at least twice the model calls per request. The author concluded that splitting tasks across agents changed who invoked the core business logic, but not what that logic does, since the deterministic boundaries remained identical in both designs.