Tokenomics: Quantifying Where Tokens Are Used in Agentic Software Engineering
5 hours ago
- #Software Engineering Automation
- #LLM-MA Systems
- #Token Consumption
- LLM-based Multi-Agent systems automate software engineering tasks, but their efficiency and resource use are not well understood.
- Analysis of token consumption in an LLM-MA system across Software Development Life Cycle stages reveals patterns.
- Code Review stage consumes the most tokens (average 59.4%), with input tokens being the largest share (average 53.9%).
- Primary cost in agentic software engineering comes from automated refinement and verification, not initial code generation.
- Findings highlight inefficiencies in agent collaboration, aiding cost prediction, workflow optimization, and future research.