Does Task Decomposition Earn Its Cost? Cost, Accuracy and Evaluation Validity of Single, Reflective and Multi-Agent LLM Analysts

This research investigates whether multi-agent task decomposition justifies its computational cost compared with single-agent and self-reflective LLM systems. Using Gemma 4 31B IT, LangGraph, and ten deterministic analytical tools, the study compares accuracy, token consumption, and latency while examining evaluation reliability and statistical power. The findings show that multi-agent orchestration substantially increases computational cost without demonstrating an accuracy advantage in the tested setting.