Researchers analyzed token usage in LLM-based multi-agent systems for software engineering, examining 30 development tasks using the ChatDev framework with a GPT-5 reasoning model. The study mapped internal phases to distinct development stages to address poorly understood operational efficiency in automated software engineering.

Findings revealed that the iterative Code Review stage accounts for 59.4% of token consumption, while input tokens make up 53.9% of total usage. This empirical evidence suggests significant inefficiencies in agentic collaboration, indicating costs lie in refinement rather than initial code generation.

The research provides a novel methodology to help practitioners predict expenses and optimize workflows. By identifying token distribution patterns, the study directs future research toward developing more token-efficient agent collaboration protocols in LLM-powered software engineering.