A new arXiv preprint reports a pre-registered study of 465 agent-authored pull requests, examining whether a code-health metric can predict the token cost of multi-agent LLM code review. The work responds to the growing practice of delegating review to LLM pipelines that bill per token.
The abstract notes that low code quality is known to raise human maintenance effort, but its effect on agent review cost has not been previously established. The study uses CodeHealth as a candidate predictor, and the title asserts that it does predict token cost, though the abstract excerpt provided does not include the full results.
Because the source is a single preprint abstract, there are no independent findings or effect sizes to compare. The significance lies in the potential to use code-quality signals to forecast LLM review expenses, which could inform budgeting and quality prioritization in organizations adopting agentic review.