On the Adoption of AI Coding Agents in Open-source Android and iOS Development

Abstract: AI coding agents are increasingly contributing to software development, yet their impact on mobile development has received little empirical attention. In this paper, we present the first category-level empirical study of agent-generated code in open-source mobile app projects. We analyzed PR acceptance behaviors across mobile platforms, agents, and task categories using 2,901 AI-authored pull requests (PRs) in 193 verified Android and iOS open-source GitHub repositories in the AIDev dataset. We find that Android projects have received 2x more AI-authored PRs and have achieved higher PR acceptance rate (71%) than iOS (63%), with significant agent-level variation on Android. Across task categories, PRs with routine tasks (feature, fix, and ui) achieve the highest acceptance, while structural changes like refactor and build achieve lower success and longer resolution times. Furthermore, our evolution analysis shows improvement in PR resolution time on Android through mid-2025 before it declined again. Our findings offer the first evidence-based characterization of AI agents effects on OSS mobile projects and establish empirical baselines for evaluating agent-generated contributions to design platform aware agentic systems.

• The paper analyzes 2,901 AI-authored pull requests across 193 Android and iOS repositories using statistical tests and validated task categories to measure acceptance and resolution patterns.
• Android PRs achieved a 71.0% acceptance rate versus 63.7% on iOS, with Codex reaching 76.8% on Android compared with 28.0% for Copilot and 42.3% for Cursor.
• Android acceptance varied sharply by task category and agent, while iOS PRs resolved about 18 times faster overall, showing that workflow infrastructure and structural tasks strongly affect agent outcomes.

Overview

This paper presents the first category-level empirical study of AI coding agent contributions to open-source mobile development. Drawing on 2,901 AI-authored pull requests (PRs) across 193 verified Android and iOS GitHub repositories, the authors characterize how PR acceptance rates and resolution times vary across mobile platforms, agents, and task categories. The study builds on the AIDev dataset (Li et al., 20 Jul 2025), which contains over 932k AI-authored PRs from five agents—Codex, Devin, Copilot, Cursor, and Claude—and extends it with additional mining from August through November 2025 using the GitHub API.

The work addresses a genuine gap: prior empirical studies of agentic coding have focused on general-purpose repositories (Watanabe et al., 18 Sep 2025) or on LLM integration as an app feature rather than agent-generated code within mobile projects (Hau et al., 21 Feb 2025). Mobile-specific studies have shown that LLMs struggle with real-world complexity—for example, APPFORGE documents end-to-end Android generation failures (Ran et al., 9 Oct 2025), and earlier work found unreliable Flutter code generation [10436306]. This paper moves from capability benchmarks to observed maintainer behavior at scale.

Dataset construction

The authors filtered the AIDev corpus to native mobile projects by selecting repositories in Java, Kotlin, Swift, or Objective-C, then applied rule-based verification of platform artifacts: AndroidManifest.xml, Gradle files, and Android directory structures for Android; Info.plist, .xcodeproj, Podfile, and structured Swift layouts for iOS. Repositories with fewer than 10 stars were removed, along with tutorials and samples identified via topic, README, and directory-structure indicators. This yielded 98 Android and 95 iOS repositories.

From these, 1,078 AI-authored PRs (728 Android, 350 iOS) came from AIDev's May–July 2025 window; extending the window to November 2025 added 1,823 more (1,626 Android, 197 iOS). The authors retain the natural Android–iOS imbalance because they claim complete coverage of OSS mobile repositories in AIDev, so the skew reflects actual differences in agent activity across ecosystems.

PR categorization used open-card sorting with GPT-5 on sampled PR titles, refined by two authors via negotiated agreement into 13 categories (feature, fix, refactor, build, chore, performance, style, test, docs, operations, ui, localization, other). Manual validation of an independent sample achieved Cohen's $\kappa = 0.877$, indicating strong inter-rater agreement. Statistical analysis relied on non-parametric tests (Mann–Whitney U, Chi-square, Kruskal–Wallis) with Holm correction, plus Bayesian add-$\alpha$ smoothing for small-sample acceptance rates.

Platform-level acceptance (RQ1)

Android projects accepted 71.0% of AI-authored PRs versus 63.7% on iOS, a statistically significant difference ($p < 0.05$). Notably, Android received roughly twice as many AI-authored PRs yet still accepted them at a higher rate—the higher submission volume did not depress acceptance.

Agent identity matters only on Android. Chi-square tests confirmed significant agent-level variation there ($p < 0.05$) but not on iOS. Codex PRs were accepted at 76.8% on Android, far above Cursor (42.3%) and Copilot (28.0%)—a nearly threefold spread between the best and worst agents. On iOS, all agents fell within a narrow 51–79% band. The practical implication is asymmetric: agent selection is a consequential decision for Android contributors but largely immaterial on iOS, where uniformly cautious review behavior appears to dominate regardless of tooling.

Task-category effects (RQ2)

On Android, category strongly predicts acceptance ($p < 0.05$). Localization reached 100% acceptance, followed by ui (88%) and fix (75%); Fisher's exact post-hoc tests showed localization significantly outperforming refactor, feature, and build. Routine, localized changes fare well, while structural modifications lag. On iOS, neither the Chi-square test nor post-hoc comparisons detected category-level variation, indicating uniform acceptance behavior across task types.

This asymmetry again carries design consequences: Android teams can target automation toward high-acceptance categories while reserving human review effort for structural changes, whereas equivalent task-specific optimization offers little expected gain on iOS.

Resolution-time dynamics (RQ3)

Resolution time exhibits a counterintuitive platform reversal relative to acceptance: despite lower acceptance rates, iOS PRs resolve approximately $18\times$ faster than Android PRs ($p < 0.05$). Codex is the fastest agent on both platforms—resolving PRs $3\times$ faster than Claude on Android—with Devin appearing fastest on iOS but excluded from reliable inference due to a sample size of only seven.

Splitting categories into functional (feature, ui, localization, fix) and non-functional groups reveals large gaps: functional PRs resolve $400\times$ faster than non-functional ones on Android and $7\times$ faster on iOS. Temporally, Android shows improvement in resolution time from May through August 2025 before a partial reversal in later months—a non-monotonic pattern confirmed by Dunn's post-hoc tests. iOS shows statistically detectable month-to-month variation but without systematic improvement. The Android decline after mid-2025 is left unexplained; it may reflect changing PR composition, reviewer fatigue, or shifts in agent behavior, but the paper does not disentangle these.

A manual inspection of ten rejected PRs surfaced rejection causes orthogonal to category or agent: rebase conflicts, request-limit errors, agent-side execution failures, failing CI pipelines, missing reviewer feedback, and auto-closure from inactivity. These suggest that infrastructure and workflow friction, not code quality alone, materially gate agent contributions.

Limitations

The authors acknowledge several constraints. Acceptance rate and resolution time are proxies that do not capture code quality, review depth, or long-term reliability. The taxonomy derives from PR titles classified by GPT-5 at default temperature, which may misrepresent mixed-intent tasks despite expert validation. Findings are bounded to active GitHub OSS projects and may not transfer to different review cultures or CI practices. Small samples for some agents and categories—most acutely Devin's seven iOS PRs—limit statistical robustness even with smoothing and corrected non-parametric tests. The causal drivers behind the observed platform differences (e.g., whether iOS's stricter design guidelines explain its lower acceptance) remain hypothesized rather than established.

Conclusion

This study establishes the first empirical baselines for AI coding agent activity in open-source mobile development, covering 2,901 PRs across 193 Android and iOS repositories. Its central findings are the divergence between the platforms—Android's higher acceptance coupled with strong sensitivity to agent choice and task category, versus iOS's lower but uniform and faster-resolution behavior—and the persistent penalty for structural changes such as refactoring and build work relative to routine fixes. The temporal analysis adds nuance: efficiency gains on Android through mid-2025 partially reversed afterward, cautioning against assuming monotonic improvement in agent–maintainer workflows. The results provide a foundation for platform-aware agentic systems, though whether the patterns generalize to cross-platform frameworks and industrial settings remains open.