AI BriefingTokenomics: Quantifying Where Tokens Go in Agentic Software Engineering
Key point
A study demonstrating that the code review stage consumes the most tokens when LLM-based multi-agent systems perform software development.
Details
This study empirically analyzed the token consumption patterns that occur when LLM-based multi-agent (LLM-MA) systems perform software development (SDLC).
Analyzing 30 software development tasks through the ChatDev framework, it was confirmed that the Code Review stage consumes the most resources, using an average of 59.4% of tokens.
The overall token composition was Input (53.9%), Output (24.4%), and Reasoning (21.6%), in that order. This suggests that the 'communication tax' arising from repeated context-passing between agents is a major driver of system operating costs.
Distinct token usage patterns exist across development stages.
- Coding stage: High share of output tokens at 58.0%
- Documentation stage: Very high share of input tokens at 80.2%
These findings provide important grounds for cost prediction and designing efficient collaboration protocols for workflow optimization in agent-based development systems.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Google Researchers Release 'Procedural Graph' for LLM Agents… Self-Evolving via Failure Trajectory AnalysisDeepSeek Releases DSec: Elastic Sandbox Infrastructure for Large-Scale Agentic Training