Researchers have conducted the first systematic analysis of token consumption in agentic software engineering systems, revealing that iterative code review accounts for 59.4% of all tokens consumed. Published on arXiv January 20, 2026 by Mohamad Salim and colleagues for presentation at MSR '26 (April 13-14 in Rio de Janeiro), the research reached 134 points on Hacker News with 57 comments on June 7, 2026.
Primary Cost Lies in Automated Refinement, Not Initial Generation
The research findings challenge assumptions about where computational costs occur in agentic systems. The primary cost of agentic software engineering lies not in initial code generation but in automated refinement and verification. Specifically, the iterative Code Review stage accounts for the majority of token consumption at an average of 59.4% of tokens across the software development lifecycle.
The researchers analyzed token consumption patterns in an LLM-MA (Multi-Agent) system within the SDLC, tracking distinct activities: requirements analysis, initial code generation, iterative code review, testing, and deployment. This novel methodology can help practitioners predict expenses and optimize workflows.
Input Tokens Constitute 53.9% of Consumption, Revealing Inefficiencies
Input tokens consistently constitute the largest share of consumption for an average of 53.9%, providing empirical evidence for potentially significant inefficiencies in agentic collaboration. This input-output asymmetry suggests agents are reading far more than they're writing, possibly re-processing context unnecessarily between agent handoffs.
The findings direct future research toward developing more token-efficient agent collaboration protocols. Specific opportunities include reducing redundant context passing between agents, caching or compressing shared context, minimizing review iterations through better initial generation, and implementing asymmetric pricing models that account for input-output imbalance.
Research Enables Cost Prediction and Architecture Optimization
For companies deploying agentic SDLC systems, this research enables several practical improvements. Organizations can now predict costs based on workflow stage distribution, identify optimization opportunities focused on the Code Review stage, allocate budgets aligned with actual token consumption patterns, and make architecture decisions that minimize expensive iterative review cycles.
The research comes at a critical time as harness engineering and fully agentic development workflows become mainstream. Understanding token economics is essential for sustainable deployment at scale, particularly as organizations move from experimental to production deployments of agentic systems.
Key Takeaways
- Iterative code review accounts for 59.4% of all token consumption in agentic software engineering systems, far exceeding initial code generation costs
- Input tokens constitute 53.9% of total consumption, suggesting significant inefficiencies in how agents pass context between stages
- The research provides the first systematic methodology for tracking and predicting token costs across the software development lifecycle
- Primary optimization opportunities lie in reducing redundant context passing and minimizing review iterations
- The findings were published January 20, 2026 on arXiv and will be presented at MSR '26 in Rio de Janeiro on April 13-14