Tokenomics: Quantifying Where Tokens Are Used in Agentic Software Engineering (arxiv.org)

A recent study has unveiled significant insights into the token usage patterns of Large Language Model-based Multi-Agent (LLM-MA) systems in software engineering, specifically within the Software Development Life Cycle (SDLC). Conducted with the ChatDev framework and powered by the GPT-5 model, the research meticulously analyzed 30 software development tasks, illustrating that the Code Review phase consumes the highest share of tokens, averaging 59.4%. Notably, input tokens dominate this consumption at 53.9%, revealing potential inefficiencies in the automation processes involved in software refinement and verification. This research is particularly noteworthy for the AI/ML community as it addresses a critical gap in understanding the operational costs and resource efficiency of agentic software engineering. By highlighting that the biggest expenses arise not from initial code creation but from subsequent automated tasks, the findings suggest new avenues for optimizing workflows. The proposed standardized evaluation framework not only aids practitioners in forecasting costs but also sets the stage for developing more efficient agent collaboration protocols in future research, thereby driving innovation in the automation of software engineering tasks.