Clean code doesn't boost AI agent pass rates but cuts tokens 8%

A new paper from Priyansh Trivedi and Olivier Schmitt at SonarSource investigates a practical question: does code cleanliness impact how AI coding agents perform? The team built minimal-pair repositories—identical in architecture, dependencies, and external behavior but differing in static-analysis rule violations and cognitive complexity. They created 33 tasks across six such pairs and ran 660 trials using Anthropic's Claude Code. The result: pass rates were unchanged between clean and messy code, but operational efficiency shifted significantly. Agents working on cleaner code used 7–8% fewer tokens and reduced file revisits by 34%, suggesting structural quality affects an agent's navigational cost without hurting task completion.

The study, submitted to arXiv (2605.20049), positions code cleanliness alongside model choice, harness design, and prompting as a factor that shapes agent behavior. By isolating cleanliness via controlled pairs that can be either degraded or cleaned, the authors offer a reproducible benchmark for future research. Their conclusion: “traditional maintainability principles remain highly relevant in the era of AI-driven development.” For teams deploying autonomous coding agents, this means investing in clean code infrastructure now pays dividends in reduced compute and faster agent iteration—even if success rates don't budge.