Tokenomics: Quantifying Where Tokens Are Used in Agentic Software Engineering

Researchers analyzed token consumption patterns in LLM-based multi-agent systems used for software engineering tasks, finding that the Code Review stage accounts for 59.4% of tokens consumed, with input tokens representing 53.9% of overall consumption. The study reveals that the primary cost of agentic software engineering comes from automated refinement and verification rather than initial code generation, highlighting inefficiencies in agent collaboration. The findings provide a framework for practitioners to predict expenses and optimize workflows while directing future research toward more token-efficient agent protocols.