AI coding agents work faster on cleaner code, but pass rates unchanged

A controlled study of autonomous coding agents examined whether the structural quality of a codebase influences performance. Using 33 tasks across six minimal‑pair repositories that differed only in static‑analysis violations and cognitive complexity, researchers found that code cleanliness did not affect the agents' pass rates. However, agents operating on cleaner code consumed 7‑8 % fewer tokens and reduced file revisitations by about 34 %, indicating lower computational overhead.

Separately, an analysis of AI coding autonomy outlines five progressive levels—from basic autocomplete (Level 0) to a fully autonomous "dark factory" (Level 5). It argues that intermediate levels, especially Level 3 where AI handles most coding tasks under human validation, offer the best balance of productivity and reliability. The report stresses the need for structured workflows, continuous feedback loops, and gradual escalation of AI capabilities to avoid challenges in reliability, transparency, and evaluation gaming at higher autonomy levels.