Agent-written infrastructure PRs are four times larger, and saying so costs you the merge
A study of 349,156 infrastructure-as-code pull requests finds agent PRs much bigger and more likely to touch core modules. Agents that disclose themselves get merged 8.4 points less often for the same kind of change, while their Terraform is no smellier than a human's.
Amine Barrak, Emna Ksontini, and Charles Wehbe look at the part of the codebase where a bad merge is most expensive: infrastructure as code, the configuration that defines cloud resources, networking, and access control. They study 349,156 IaC pull requests from 49,241 GitHub repositories and sort authors into four groups: humans, automation bots, agents that leave traces but don't announce themselves ("silent" agents), and agents that disclose they wrote the change.
The first finding is about the denominator. 32.9% of pull requests that look human actually come from bots, and 12.3% of the remainder carry agent traces, a share that grows over the study period. Agent PRs are about four times larger than human ones and more often modify core infrastructure modules, while bots mostly bump dependencies.
Merge rates split by whether the agent says what it is. Silent-agent PRs merge more often overall than human PRs (87.4% vs. 78.9%), but 4.0 points less often when compared inside the same repositories. Disclosed agents create new infrastructure more often than humans (34.0% vs. 21.4%), and even for the same change type in the same repository their PRs are 8.4 points less likely to be merged. On quality, 2,645 matched agent-human Terraform pairs introduce sustainability smells at nearly the same rate (10.4% vs. 10.3%), most often missing remote state and monolithic configurations.
Why it matters for code review
Reviewers appear to be grading the label, not the diff. Matched Terraform changes are about as clean whoever wrote them, yet the disclosed ones lose merges. That turns disclosure into a tax, which pushes teams toward the silent path and makes review policy harder to apply to the very changes that most need it.
The size finding is the other warning for anyone running agents around the clock against infrastructure repos. Four-times-larger PRs that touch core modules are exactly the changes a human can't read well at volume. The paper's own numbers suggest the fix is a policy that routes by what a change touches and how big it is, rather than by who claims authorship.