A framework shows where AI infrastructure changes slow merges

When an AI agent changes both application code and deployment setup, a passing local test does not tell us whether the contribution will move smoothly through a shared repository.

It remains unclear whether the extra infrastructure change creates integration friction, and whether human maintainers inspect agent-authored work consistently.

The researchers classify file roles and match comparable pull requests to measure how co-changing application code and infrastructure affects merge time and human review.

Researchers classified every modified file with ordered rules, then matched dual-scope and application-only pull requests within the same agent architecture, repository type, test-file status, language ecosystem, and similar application-file volume.

This reduces measured differences caused by work size or context, though it cannot remove unobserved intent or prompting variation.

In AIDev v5, dual-scope requests had a 16.60-hour median merge latency versus 3.47 hours for controls, and the model estimated 3.237 times the control's total merge time. The MOSAIC-3M comparison showed no significant delay, linking the answer to the repository cohort and its integration process.

The comparison is observational and covers pull requests authored from June 1 through July 31, 2025, in mature repositories with more than 100 stars; evaluated requests had to have a terminal merge timestamp by August 31, 2025.

Abandoned requests are unseen, and matching excludes extreme file-volume cases, so the latency estimates are conservative lower bounds for typical matched work rather than a universal causal effect.