Code Cleanliness Does Not Affect AI Coding Agent Performance

A new study from arXiv (2605.20049) investigates whether code cleanliness impacts the performance of autonomous coding agents. The researchers developed an evaluation protocol using minimal-pair repositories that match in architecture, dependencies, and external behavior but differ in static-analysis rule violations and cognitive complexity. They created 33 tasks across six pairs, with repositories modified in both directions—either degrading clean code or cleaning messy code. In 660 trials using Claude Code, the study found that code cleanliness does not change the agent's pass rate. The findings challenge assumptions about code quality's importance for AI agents, suggesting that agent capability may be independent of code structure.