Study Finds Agentic AI Boosts Code Size but Not Architecture Quality

A new causal study of 151 open-source Java repositories reveals that agentic AI adoption (popularly known as 'vibe coding') significantly inflates codebase size without improving architectural quality. Researchers from the paper—accepted at SEAA 2026—used a staggered difference-in-differences design on 74 repos with detectable agentic AI usage (via config files or Co-Authored-By trailers) matched to 77 controls. Over a 13-month window, they found that lines of code grew 12.8% (p=0.003) after AI adoption, while total architectural smell counts remained essentially unchanged (+1.1%, p=0.82). Architectural smell density dropped 6.7% (p=0.004), but the authors attribute this entirely to the denominator effect of larger codebases, not to any genuine architectural improvement. Pre-trends were flat, and robustness checks (wild cluster bootstrap, Lee bounds) confirmed the pattern.

The findings carry important implications for the software engineering community. Researchers caution that using density‑normalized metrics (e.g., smells per line of code) can mislead when the treatment affects system size. In this case, the apparent decline in smell density could be misinterpreted as an AI‑driven quality gain. To avoid such pitfalls, the authors recommend reporting raw counts and explicit decompositions in causal mining studies of AI tool adoption. For practitioners, the study suggests that while agentic AI accelerates code production, it does not inherently improve architectural health—and may even degrade maintainability if total smell counts remain flat while code volume grows. The full replication package, including the curated monthly panel of 151 repositories and 1,811 Arcan snapshots, is publicly available.