Current software‑engineering agents are mostly evaluated inside isolated sandboxes using synchronous local tests. Deploying new features in collaborative repositories, however, requires simultaneous changes to CI/CD pipelines, containers, provisioning manifests, and application code. Because agents lack deployment awareness, cross‑domain edits introduce unverified dependencies into asynchronous integration pipelines. To isolate the causal integration overhead of dual‑scope pull requests (modifying both application code and deployment infrastructure), we examined 26,999 PRs across the AIDev v5 and MOSAIC‑3M datasets, covering six agent architectures and an aligned human baseline. We applied a deterministic four‑tier file taxonomy and an exact 4‑dimensional strata‑matching framework, ensuring conditional exchangeability while controlling for application size, to extract integration metrics. Findings show agents act mainly as application coders or choreographers, with maintenance tasks accounting for 17.70%–25.41% of agent PRs. Dual‑scope changes represent only 4.48%–4.55% of agent PRs but produce a bimodal integration pattern. In the first mode, explicitly reviewed contributions suffer significant delays, adding a median +13.13 hours and inflating time‑to‑merge by 3.24×, driven mechanically by CI/CD and provisioning manifests. In the second mode, maintainer gatekeeping exhibits a structural bypass: even external contributors without write access see a 63.7% reduction in the odds of human review compared to human peers. Consequently, 43.1%–56.2% of dual‑scope PRs merge silently within a median of under 31 minutes without inspection. These unchecked modifications weaken build verification, highlighting the need for deployment‑aware benchmarks and strict path‑based access controls.
Review