Tokenomics: A Quantitative Analysis of Token Usage in Agent-Based Software Engineering

LLM-based multi-agent (LLM-MA) systems are being introduced into the software development lifecycle (SDLC), but unpredictable costs and resource consumption remain unresolved challenges.

This study used the ChatDev framework and the GPT-5 reasoning model to analyze token consumption patterns across each development stage, including design, coding, code review, and testing. The key findings are as follows.

• The Code Review stage consumes the most resources, accounting for an average of 59.4% of total token consumption.
• Input tokens account for an average of 53.9%, suggesting significant inefficiency in the collaboration process between agents.

In conclusion, the main costs of agent-based software engineering arise not from initial code generation but from the automated refinement and verification process. This research can contribute to optimizing future cost prediction and developing more token-efficient agent collaboration protocols.