Unhealthy code makes AI agents consume 35-50% more tokens (codescene.com)

Recent research by CodeScene reveals that AI agents working with unhealthy codebases consume 35-50% more tokens to complete tasks than those using healthier code. This study, conducted across C++, Java, and Python, demonstrates that not only does the quality of code affect performance and output, but it also significantly impacts the operational costs associated with token consumption in AI models. For instance, C++ files scored low on maintainability showed a staggering 43.8% increase in token waste during test case generation, translating to nearly twice the costs for teams working with degraded code. This finding is critical for the AI and ML community as it highlights the substantial effects of code quality on both productivity and budget management in enterprise environments transitioning to agentic AI systems. The research indicates a clear trend: healthier code directly correlates to reduced token consumption, leading to cost savings and improved AI performance. To address these challenges, CodeScene proposes the integration of a Code Health metric into AI workflows, which can safeguard new code and uplift existing codebases, thereby ensuring that enterprises optimize their coding processes and cut down on unnecessary token expenses.